{
    "schemaVersion": 1,
    "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA RTX PRO 6000 Blackwell Server Edition", "totalGlobalMem": 101975851008,
      "computeMajor": 12, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 1536,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 188
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 101376, "sharedMemPerMultiprocessor": 102400
    }
    ],
      "trace_id": "E9CCE101E22C43969E216307C4C51492",
      "profile_memory": 1,
      "cuda_driver_version": 13020,
      "record_shapes": 1,
      "with_flops": 1,
      "cuda_runtime_version": 13000,
      "cupti_version": 130001,
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1790857026000000000,
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507957376.709, "dur": 2425.801,
    "args": {
      "External id": 513,"Record function id": 0, "Sequence number": 4685, "Fwd thread id": 1, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507957430.183, "dur": 2231.680,
    "args": {
      "External id": 514,"Sequence number": 4685, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 1
    }
  },
  {
    "ph": "f", "id": 1, "pid": 14565, "tid": 14640, "ts": 1290507957430.183,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290507957554.177, "dur": 2086.008,
    "args": {
      "External id": 515,"Record function id": 0, "scope": "user_scope", "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290507958542.430, "dur": 24.379,
    "args": {
      "External id": 516,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959823.132, "dur": 48.005,
    "args": {
      "External id": 517,"Record function id": 0, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959860.661, "dur": 8.691,
    "args": {
      "External id": 518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959865.632, "dur": 2.845,
    "args": {
      "External id": 519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959866.645, "dur": 1.634,
    "args": {
      "External id": 520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959875.551, "dur": 4.425,
    "args": {
      "External id": 521,"Record function id": 0, "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959877.083, "dur": 2.155,
    "args": {
      "External id": 522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959877.861, "dur": 1.047,
    "args": {
      "External id": 523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959878.407, "dur": 0.337,
    "args": {
      "External id": 524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959884.386, "dur": 4.381,
    "args": {
      "External id": 525,"Record function id": 0, "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959885.823, "dur": 2.148,
    "args": {
      "External id": 526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959886.626, "dur": 1.030,
    "args": {
      "External id": 527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959887.170, "dur": 0.356,
    "args": {
      "External id": 528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959892.206, "dur": 3.738,
    "args": {
      "External id": 529,"Record function id": 0, "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959893.512, "dur": 1.727,
    "args": {
      "External id": 530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959894.048, "dur": 0.846,
    "args": {
      "External id": 531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959894.410, "dur": 0.342,
    "args": {
      "External id": 532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959899.262, "dur": 3.890,
    "args": {
      "External id": 533,"Record function id": 0, "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959900.496, "dur": 1.943,
    "args": {
      "External id": 534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 21
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959901.116, "dur": 0.996,
    "args": {
      "External id": 535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 22
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959901.483, "dur": 0.466,
    "args": {
      "External id": 536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 23
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959906.311, "dur": 3.932,
    "args": {
      "External id": 537,"Record function id": 0, "Ev Idx": 24
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959907.685, "dur": 1.883,
    "args": {
      "External id": 538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 25
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959908.258, "dur": 0.965,
    "args": {
      "External id": 539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959908.631, "dur": 0.427,
    "args": {
      "External id": 540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 27
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959913.460, "dur": 3.808,
    "args": {
      "External id": 541,"Record function id": 0, "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959914.753, "dur": 1.845,
    "args": {
      "External id": 542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 29
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959915.328, "dur": 0.964,
    "args": {
      "External id": 543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959915.710, "dur": 0.420,
    "args": {
      "External id": 544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 31
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959920.477, "dur": 3.748,
    "args": {
      "External id": 545,"Record function id": 0, "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959921.737, "dur": 1.815,
    "args": {
      "External id": 546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 33
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959922.355, "dur": 0.871,
    "args": {
      "External id": 547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 34
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959922.718, "dur": 0.379,
    "args": {
      "External id": 548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 35
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959927.462, "dur": 4.657,
    "args": {
      "External id": 549,"Record function id": 0, "Ev Idx": 36
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959928.679, "dur": 2.738,
    "args": {
      "External id": 550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 37
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959929.301, "dur": 1.759,
    "args": {
      "External id": 551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 38
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959930.554, "dur": 0.371,
    "args": {
      "External id": 552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 39
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959935.374, "dur": 4.279,
    "args": {
      "External id": 553,"Record function id": 0, "Ev Idx": 40
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959936.545, "dur": 2.403,
    "args": {
      "External id": 554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 41
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959937.310, "dur": 0.852,
    "args": {
      "External id": 555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 42
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959937.743, "dur": 0.288,
    "args": {
      "External id": 556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 43
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959942.837, "dur": 3.782,
    "args": {
      "External id": 557,"Record function id": 0, "Ev Idx": 44
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959943.994, "dur": 1.946,
    "args": {
      "External id": 558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 45
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959944.681, "dur": 0.826,
    "args": {
      "External id": 559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 46
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959945.080, "dur": 0.297,
    "args": {
      "External id": 560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 47
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959949.854, "dur": 3.863,
    "args": {
      "External id": 561,"Record function id": 0, "Ev Idx": 48
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959951.084, "dur": 1.949,
    "args": {
      "External id": 562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 49
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959951.691, "dur": 1.037,
    "args": {
      "External id": 563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 50
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959952.080, "dur": 0.509,
    "args": {
      "External id": 564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 51
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959956.904, "dur": 3.667,
    "args": {
      "External id": 565,"Record function id": 0, "Ev Idx": 52
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959958.151, "dur": 1.747,
    "args": {
      "External id": 566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 53
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959958.714, "dur": 0.875,
    "args": {
      "External id": 567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 54
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959959.088, "dur": 0.358,
    "args": {
      "External id": 568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 55
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959963.751, "dur": 3.821,
    "args": {
      "External id": 569,"Record function id": 0, "Ev Idx": 56
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959965.036, "dur": 1.864,
    "args": {
      "External id": 570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 57
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959965.645, "dur": 0.916,
    "args": {
      "External id": 571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 58
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959966.050, "dur": 0.364,
    "args": {
      "External id": 572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 59
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959970.783, "dur": 3.643,
    "args": {
      "External id": 573,"Record function id": 0, "Ev Idx": 60
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959971.936, "dur": 1.829,
    "args": {
      "External id": 574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 61
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959972.623, "dur": 0.833,
    "args": {
      "External id": 575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 62
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959972.997, "dur": 0.333,
    "args": {
      "External id": 576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 63
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959977.558, "dur": 4.788,
    "args": {
      "External id": 577,"Record function id": 0, "Ev Idx": 64
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959978.832, "dur": 2.805,
    "args": {
      "External id": 578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 65
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959980.218, "dur": 1.094,
    "args": {
      "External id": 579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 66
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959980.817, "dur": 0.351,
    "args": {
      "External id": 580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 67
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959985.559, "dur": 4.004,
    "args": {
      "External id": 581,"Record function id": 0, "Ev Idx": 68
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959986.876, "dur": 1.976,
    "args": {
      "External id": 582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 69
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959987.565, "dur": 0.986,
    "args": {
      "External id": 583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 70
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959988.047, "dur": 0.351,
    "args": {
      "External id": 584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 71
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959992.746, "dur": 3.674,
    "args": {
      "External id": 585,"Record function id": 0, "Ev Idx": 72
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959993.930, "dur": 1.790,
    "args": {
      "External id": 586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 73
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959994.562, "dur": 0.853,
    "args": {
      "External id": 587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 74
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507959994.933, "dur": 0.342,
    "args": {
      "External id": 588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 75
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507959999.595, "dur": 4.554,
    "args": {
      "External id": 589,"Record function id": 0, "Ev Idx": 76
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960001.637, "dur": 1.810,
    "args": {
      "External id": 590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 77
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960002.251, "dur": 0.899,
    "args": {
      "External id": 591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 78
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960002.612, "dur": 0.395,
    "args": {
      "External id": 592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 79
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960007.427, "dur": 3.889,
    "args": {
      "External id": 593,"Record function id": 0, "Ev Idx": 80
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960008.827, "dur": 1.800,
    "args": {
      "External id": 594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 81
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960009.437, "dur": 0.888,
    "args": {
      "External id": 595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 82
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960009.807, "dur": 0.368,
    "args": {
      "External id": 596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 83
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960014.715, "dur": 3.977,
    "args": {
      "External id": 597,"Record function id": 0, "Ev Idx": 84
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960015.993, "dur": 1.992,
    "args": {
      "External id": 598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 85
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960016.683, "dur": 0.893,
    "args": {
      "External id": 599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 86
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960017.085, "dur": 0.343,
    "args": {
      "External id": 600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 87
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960021.979, "dur": 4.230,
    "args": {
      "External id": 601,"Record function id": 0, "Ev Idx": 88
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960023.145, "dur": 2.354,
    "args": {
      "External id": 602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 89
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960023.875, "dur": 1.204,
    "args": {
      "External id": 603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 90
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960024.240, "dur": 0.690,
    "args": {
      "External id": 604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 91
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960029.523, "dur": 36.773,
    "args": {
      "External id": 605,"Record function id": 0, "Ev Idx": 92
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960031.612, "dur": 33.346,
    "args": {
      "External id": 606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 93
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960032.206, "dur": 32.070,
    "args": {
      "External id": 607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 94
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960032.617, "dur": 30.345,
    "args": {
      "External id": 608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 95
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960074.255, "dur": 9.667,
    "args": {
      "External id": 609,"Record function id": 0, "Ev Idx": 96
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960077.430, "dur": 2.036,
    "args": {
      "External id": 610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 97
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960078.216, "dur": 0.935,
    "args": {
      "External id": 611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 98
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960078.665, "dur": 0.346,
    "args": {
      "External id": 612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 99
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960087.212, "dur": 3.868,
    "args": {
      "External id": 613,"Record function id": 0, "Ev Idx": 100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960088.621, "dur": 1.762,
    "args": {
      "External id": 614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960089.209, "dur": 0.847,
    "args": {
      "External id": 615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960089.599, "dur": 0.316,
    "args": {
      "External id": 616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960094.289, "dur": 4.054,
    "args": {
      "External id": 617,"Record function id": 0, "Ev Idx": 104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960095.649, "dur": 1.936,
    "args": {
      "External id": 618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960096.273, "dur": 1.005,
    "args": {
      "External id": 619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960096.751, "dur": 0.382,
    "args": {
      "External id": 620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960101.523, "dur": 3.903,
    "args": {
      "External id": 621,"Record function id": 0, "Ev Idx": 108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960102.733, "dur": 1.947,
    "args": {
      "External id": 622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960103.376, "dur": 1.015,
    "args": {
      "External id": 623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960103.949, "dur": 0.311,
    "args": {
      "External id": 624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960108.554, "dur": 3.782,
    "args": {
      "External id": 625,"Record function id": 0, "Ev Idx": 112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960109.775, "dur": 1.863,
    "args": {
      "External id": 626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960110.335, "dur": 0.989,
    "args": {
      "External id": 627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960110.796, "dur": 0.383,
    "args": {
      "External id": 628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960115.435, "dur": 3.715,
    "args": {
      "External id": 629,"Record function id": 0, "Ev Idx": 116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960116.622, "dur": 1.815,
    "args": {
      "External id": 630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960117.232, "dur": 0.905,
    "args": {
      "External id": 631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960117.637, "dur": 0.351,
    "args": {
      "External id": 632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960123.560, "dur": 4.038,
    "args": {
      "External id": 633,"Record function id": 0, "Ev Idx": 120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960124.844, "dur": 2.062,
    "args": {
      "External id": 634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960125.538, "dur": 1.054,
    "args": {
      "External id": 635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960126.034, "dur": 0.387,
    "args": {
      "External id": 636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960130.730, "dur": 3.792,
    "args": {
      "External id": 637,"Record function id": 0, "Ev Idx": 124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960132.028, "dur": 1.804,
    "args": {
      "External id": 638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960132.645, "dur": 0.874,
    "args": {
      "External id": 639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960133.018, "dur": 0.356,
    "args": {
      "External id": 640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960137.640, "dur": 4.023,
    "args": {
      "External id": 641,"Record function id": 0, "Ev Idx": 128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960138.878, "dur": 2.019,
    "args": {
      "External id": 642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960139.575, "dur": 0.807,
    "args": {
      "External id": 643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960139.942, "dur": 0.308,
    "args": {
      "External id": 644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960144.872, "dur": 3.811,
    "args": {
      "External id": 645,"Record function id": 0, "Ev Idx": 132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960146.032, "dur": 1.890,
    "args": {
      "External id": 646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960146.720, "dur": 0.784,
    "args": {
      "External id": 647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960147.080, "dur": 0.297,
    "args": {
      "External id": 648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960151.850, "dur": 3.698,
    "args": {
      "External id": 649,"Record function id": 0, "Ev Idx": 136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960153.138, "dur": 1.719,
    "args": {
      "External id": 650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960153.706, "dur": 0.845,
    "args": {
      "External id": 651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960154.080, "dur": 0.330,
    "args": {
      "External id": 652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960158.683, "dur": 3.588,
    "args": {
      "External id": 653,"Record function id": 0, "Ev Idx": 140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960159.825, "dur": 1.755,
    "args": {
      "External id": 654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960160.410, "dur": 0.860,
    "args": {
      "External id": 655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960160.769, "dur": 0.362,
    "args": {
      "External id": 656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960165.466, "dur": 5.171,
    "args": {
      "External id": 657,"Record function id": 0, "Ev Idx": 144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960166.739, "dur": 3.185,
    "args": {
      "External id": 658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960167.302, "dur": 2.275,
    "args": {
      "External id": 659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960168.464, "dur": 0.951,
    "args": {
      "External id": 660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960173.834, "dur": 3.765,
    "args": {
      "External id": 661,"Record function id": 0, "Ev Idx": 148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960175.122, "dur": 1.767,
    "args": {
      "External id": 662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960175.715, "dur": 0.866,
    "args": {
      "External id": 663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960176.100, "dur": 0.345,
    "args": {
      "External id": 664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960180.866, "dur": 3.912,
    "args": {
      "External id": 665,"Record function id": 0, "Ev Idx": 152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960182.222, "dur": 1.866,
    "args": {
      "External id": 666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960182.815, "dur": 0.968,
    "args": {
      "External id": 667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960183.191, "dur": 0.442,
    "args": {
      "External id": 668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960187.940, "dur": 3.845,
    "args": {
      "External id": 669,"Record function id": 0, "Ev Idx": 156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960189.276, "dur": 1.806,
    "args": {
      "External id": 670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960189.865, "dur": 0.900,
    "args": {
      "External id": 671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960190.238, "dur": 0.380,
    "args": {
      "External id": 672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960194.906, "dur": 4.016,
    "args": {
      "External id": 673,"Record function id": 0, "Ev Idx": 160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960196.098, "dur": 2.089,
    "args": {
      "External id": 674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960196.660, "dur": 1.221,
    "args": {
      "External id": 675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960197.021, "dur": 0.714,
    "args": {
      "External id": 676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960202.053, "dur": 3.965,
    "args": {
      "External id": 677,"Record function id": 0, "Ev Idx": 164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960203.371, "dur": 1.959,
    "args": {
      "External id": 678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960203.930, "dur": 1.090,
    "args": {
      "External id": 679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960204.301, "dur": 0.573,
    "args": {
      "External id": 680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960209.194, "dur": 3.885,
    "args": {
      "External id": 681,"Record function id": 0, "Ev Idx": 168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960210.443, "dur": 1.946,
    "args": {
      "External id": 682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960211.011, "dur": 1.060,
    "args": {
      "External id": 683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960211.372, "dur": 0.560,
    "args": {
      "External id": 684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960216.571, "dur": 39.180,
    "args": {
      "External id": 685,"Record function id": 0, "Ev Idx": 172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960217.805, "dur": 35.916,
    "args": {
      "External id": 686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960219.751, "dur": 33.078,
    "args": {
      "External id": 687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960220.159, "dur": 31.421,
    "args": {
      "External id": 688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960263.068, "dur": 6.028,
    "args": {
      "External id": 689,"Record function id": 0, "Ev Idx": 176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960266.068, "dur": 2.283,
    "args": {
      "External id": 690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960266.990, "dur": 0.920,
    "args": {
      "External id": 691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960267.423, "dur": 0.342,
    "args": {
      "External id": 692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960272.436, "dur": 3.824,
    "args": {
      "External id": 693,"Record function id": 0, "Ev Idx": 180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960273.762, "dur": 1.775,
    "args": {
      "External id": 694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960274.402, "dur": 0.830,
    "args": {
      "External id": 695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960274.787, "dur": 0.314,
    "args": {
      "External id": 696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960279.442, "dur": 3.566,
    "args": {
      "External id": 697,"Record function id": 0, "Ev Idx": 184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960280.634, "dur": 1.670,
    "args": {
      "External id": 698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960281.186, "dur": 0.808,
    "args": {
      "External id": 699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960281.544, "dur": 0.320,
    "args": {
      "External id": 700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960286.097, "dur": 3.655,
    "args": {
      "External id": 701,"Record function id": 0, "Ev Idx": 188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960287.331, "dur": 1.725,
    "args": {
      "External id": 702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960287.890, "dur": 0.858,
    "args": {
      "External id": 703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960288.299, "dur": 0.317,
    "args": {
      "External id": 704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960292.828, "dur": 3.649,
    "args": {
      "External id": 705,"Record function id": 0, "Ev Idx": 192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960294.050, "dur": 1.712,
    "args": {
      "External id": 706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960294.603, "dur": 0.852,
    "args": {
      "External id": 707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960295.010, "dur": 0.314,
    "args": {
      "External id": 708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960299.576, "dur": 3.634,
    "args": {
      "External id": 709,"Record function id": 0, "Ev Idx": 196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960300.762, "dur": 1.753,
    "args": {
      "External id": 710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960301.362, "dur": 0.840,
    "args": {
      "External id": 711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960301.757, "dur": 0.310,
    "args": {
      "External id": 712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960306.306, "dur": 3.673,
    "args": {
      "External id": 713,"Record function id": 0, "Ev Idx": 200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960307.525, "dur": 1.757,
    "args": {
      "External id": 714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960308.127, "dur": 0.840,
    "args": {
      "External id": 715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960308.517, "dur": 0.312,
    "args": {
      "External id": 716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960313.104, "dur": 3.583,
    "args": {
      "External id": 717,"Record function id": 0, "Ev Idx": 204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960314.264, "dur": 1.724,
    "args": {
      "External id": 718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960314.851, "dur": 0.825,
    "args": {
      "External id": 719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960315.236, "dur": 0.311,
    "args": {
      "External id": 720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960319.781, "dur": 3.664,
    "args": {
      "External id": 721,"Record function id": 0, "Ev Idx": 208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960321.023, "dur": 1.728,
    "args": {
      "External id": 722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960321.617, "dur": 0.821,
    "args": {
      "External id": 723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960321.981, "dur": 0.316,
    "args": {
      "External id": 724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960326.572, "dur": 3.715,
    "args": {
      "External id": 725,"Record function id": 0, "Ev Idx": 212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960327.857, "dur": 1.725,
    "args": {
      "External id": 726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960328.432, "dur": 0.833,
    "args": {
      "External id": 727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960328.817, "dur": 0.318,
    "args": {
      "External id": 728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960333.394, "dur": 3.961,
    "args": {
      "External id": 729,"Record function id": 0, "Ev Idx": 216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960334.651, "dur": 1.981,
    "args": {
      "External id": 730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960335.427, "dur": 0.799,
    "args": {
      "External id": 731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960335.796, "dur": 0.303,
    "args": {
      "External id": 732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960340.470, "dur": 3.738,
    "args": {
      "External id": 733,"Record function id": 0, "Ev Idx": 220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960341.610, "dur": 1.886,
    "args": {
      "External id": 734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960342.288, "dur": 0.800,
    "args": {
      "External id": 735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960342.646, "dur": 0.315,
    "args": {
      "External id": 736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960347.298, "dur": 4.177,
    "args": {
      "External id": 737,"Record function id": 0, "Ev Idx": 224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960348.558, "dur": 2.200,
    "args": {
      "External id": 738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960349.120, "dur": 1.328,
    "args": {
      "External id": 739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960349.498, "dur": 0.806,
    "args": {
      "External id": 740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960354.737, "dur": 3.773,
    "args": {
      "External id": 741,"Record function id": 0, "Ev Idx": 228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960355.991, "dur": 1.803,
    "args": {
      "External id": 742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960356.576, "dur": 0.910,
    "args": {
      "External id": 743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960356.955, "dur": 0.383,
    "args": {
      "External id": 744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960361.694, "dur": 3.807,
    "args": {
      "External id": 745,"Record function id": 0, "Ev Idx": 232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960363.024, "dur": 1.780,
    "args": {
      "External id": 746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960363.614, "dur": 0.879,
    "args": {
      "External id": 747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960363.998, "dur": 0.357,
    "args": {
      "External id": 748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960368.655, "dur": 3.570,
    "args": {
      "External id": 749,"Record function id": 0, "Ev Idx": 236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960369.842, "dur": 1.686,
    "args": {
      "External id": 750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960370.391, "dur": 0.834,
    "args": {
      "External id": 751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960370.760, "dur": 0.328,
    "args": {
      "External id": 752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960375.328, "dur": 3.917,
    "args": {
      "External id": 753,"Record function id": 0, "Ev Idx": 240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960376.508, "dur": 2.047,
    "args": {
      "External id": 754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960377.071, "dur": 1.158,
    "args": {
      "External id": 755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960377.445, "dur": 0.638,
    "args": {
      "External id": 756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960382.381, "dur": 3.906,
    "args": {
      "External id": 757,"Record function id": 0, "Ev Idx": 244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960383.545, "dur": 2.015,
    "args": {
      "External id": 758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960384.107, "dur": 1.138,
    "args": {
      "External id": 759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960384.500, "dur": 0.595,
    "args": {
      "External id": 760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960389.507, "dur": 3.802,
    "args": {
      "External id": 761,"Record function id": 0, "Ev Idx": 248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960390.863, "dur": 1.775,
    "args": {
      "External id": 762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960391.412, "dur": 0.915,
    "args": {
      "External id": 763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960391.792, "dur": 0.380,
    "args": {
      "External id": 764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960396.674, "dur": 3.952,
    "args": {
      "External id": 765,"Record function id": 0, "Ev Idx": 252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960397.960, "dur": 1.985,
    "args": {
      "External id": 766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960398.526, "dur": 1.111,
    "args": {
      "External id": 767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960398.902, "dur": 0.587,
    "args": {
      "External id": 768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960403.793, "dur": 34.025,
    "args": {
      "External id": 769,"Record function id": 0, "Ev Idx": 256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960405.079, "dur": 31.408,
    "args": {
      "External id": 770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960405.661, "dur": 30.150,
    "args": {
      "External id": 771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960406.038, "dur": 28.594,
    "args": {
      "External id": 772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960445.019, "dur": 6.205,
    "args": {
      "External id": 773,"Record function id": 0, "Ev Idx": 260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960448.037, "dur": 2.423,
    "args": {
      "External id": 774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960448.983, "dur": 0.948,
    "args": {
      "External id": 775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960449.407, "dur": 0.367,
    "args": {
      "External id": 776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960454.531, "dur": 4.005,
    "args": {
      "External id": 777,"Record function id": 0, "Ev Idx": 264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960455.790, "dur": 1.998,
    "args": {
      "External id": 778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960456.568, "dur": 0.807,
    "args": {
      "External id": 779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960456.910, "dur": 0.331,
    "args": {
      "External id": 780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960461.695, "dur": 3.786,
    "args": {
      "External id": 781,"Record function id": 0, "Ev Idx": 268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960462.942, "dur": 1.821,
    "args": {
      "External id": 782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960463.605, "dur": 0.856,
    "args": {
      "External id": 783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960464.027, "dur": 0.302,
    "args": {
      "External id": 784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960468.641, "dur": 3.674,
    "args": {
      "External id": 785,"Record function id": 0, "Ev Idx": 272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960469.955, "dur": 1.654,
    "args": {
      "External id": 786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960470.503, "dur": 0.786,
    "args": {
      "External id": 787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960470.836, "dur": 0.326,
    "args": {
      "External id": 788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960475.424, "dur": 3.685,
    "args": {
      "External id": 789,"Record function id": 0, "Ev Idx": 276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960476.672, "dur": 1.741,
    "args": {
      "External id": 790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960477.234, "dur": 0.869,
    "args": {
      "External id": 791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960477.615, "dur": 0.344,
    "args": {
      "External id": 792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960482.192, "dur": 3.640,
    "args": {
      "External id": 793,"Record function id": 0, "Ev Idx": 280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960483.466, "dur": 1.681,
    "args": {
      "External id": 794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960484.023, "dur": 0.814,
    "args": {
      "External id": 795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960484.382, "dur": 0.326,
    "args": {
      "External id": 796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960488.934, "dur": 3.637,
    "args": {
      "External id": 797,"Record function id": 0, "Ev Idx": 284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960490.189, "dur": 1.698,
    "args": {
      "External id": 798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960490.742, "dur": 0.831,
    "args": {
      "External id": 799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960491.117, "dur": 0.328,
    "args": {
      "External id": 800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960495.634, "dur": 3.662,
    "args": {
      "External id": 801,"Record function id": 0, "Ev Idx": 288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960496.868, "dur": 1.736,
    "args": {
      "External id": 802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960497.432, "dur": 0.853,
    "args": {
      "External id": 803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960497.834, "dur": 0.323,
    "args": {
      "External id": 804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960502.411, "dur": 3.840,
    "args": {
      "External id": 805,"Record function id": 0, "Ev Idx": 292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960503.581, "dur": 1.930,
    "args": {
      "External id": 806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960504.249, "dur": 0.842,
    "args": {
      "External id": 807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960504.620, "dur": 0.333,
    "args": {
      "External id": 808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960509.344, "dur": 3.866,
    "args": {
      "External id": 809,"Record function id": 0, "Ev Idx": 296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960510.514, "dur": 1.983,
    "args": {
      "External id": 810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960511.305, "dur": 0.779,
    "args": {
      "External id": 811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960511.642, "dur": 0.311,
    "args": {
      "External id": 812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960516.330, "dur": 3.496,
    "args": {
      "External id": 813,"Record function id": 0, "Ev Idx": 300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960517.477, "dur": 1.651,
    "args": {
      "External id": 814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960518.017, "dur": 0.806,
    "args": {
      "External id": 815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960518.354, "dur": 0.331,
    "args": {
      "External id": 816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960522.916, "dur": 3.843,
    "args": {
      "External id": 817,"Record function id": 0, "Ev Idx": 304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960524.085, "dur": 1.910,
    "args": {
      "External id": 818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960524.761, "dur": 0.818,
    "args": {
      "External id": 819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960525.119, "dur": 0.329,
    "args": {
      "External id": 820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960529.869, "dur": 4.191,
    "args": {
      "External id": 821,"Record function id": 0, "Ev Idx": 308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960530.998, "dur": 2.357,
    "args": {
      "External id": 822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960531.663, "dur": 1.379,
    "args": {
      "External id": 823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960532.077, "dur": 0.837,
    "args": {
      "External id": 824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960537.216, "dur": 3.675,
    "args": {
      "External id": 825,"Record function id": 0, "Ev Idx": 312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960538.438, "dur": 1.725,
    "args": {
      "External id": 826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960539.039, "dur": 0.846,
    "args": {
      "External id": 827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960539.384, "dur": 0.355,
    "args": {
      "External id": 828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960544.109, "dur": 3.783,
    "args": {
      "External id": 829,"Record function id": 0, "Ev Idx": 316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960545.431, "dur": 1.754,
    "args": {
      "External id": 830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960546.004, "dur": 0.887,
    "args": {
      "External id": 831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960546.396, "dur": 0.365,
    "args": {
      "External id": 832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960550.981, "dur": 3.782,
    "args": {
      "External id": 833,"Record function id": 0, "Ev Idx": 320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960552.283, "dur": 1.770,
    "args": {
      "External id": 834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960552.830, "dur": 0.903,
    "args": {
      "External id": 835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960553.200, "dur": 0.371,
    "args": {
      "External id": 836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960557.921, "dur": 4.134,
    "args": {
      "External id": 837,"Record function id": 0, "Ev Idx": 324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960559.217, "dur": 2.125,
    "args": {
      "External id": 838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960559.770, "dur": 1.255,
    "args": {
      "External id": 839,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960560.219, "dur": 0.669,
    "args": {
      "External id": 840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960565.244, "dur": 4.036,
    "args": {
      "External id": 841,"Record function id": 0, "Ev Idx": 328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960566.530, "dur": 1.998,
    "args": {
      "External id": 842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960567.139, "dur": 1.109,
    "args": {
      "External id": 843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960567.520, "dur": 0.593,
    "args": {
      "External id": 844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960572.479, "dur": 3.910,
    "args": {
      "External id": 845,"Record function id": 0, "Ev Idx": 332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960573.855, "dur": 1.823,
    "args": {
      "External id": 846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960574.452, "dur": 0.938,
    "args": {
      "External id": 847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960574.887, "dur": 0.358,
    "args": {
      "External id": 848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960579.571, "dur": 3.860,
    "args": {
      "External id": 849,"Record function id": 0, "Ev Idx": 336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960580.819, "dur": 1.894,
    "args": {
      "External id": 850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960581.405, "dur": 1.002,
    "args": {
      "External id": 851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960581.784, "dur": 0.491,
    "args": {
      "External id": 852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960586.634, "dur": 82.262,
    "args": {
      "External id": 853,"Record function id": 0, "Ev Idx": 340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960587.873, "dur": 75.382,
    "args": {
      "External id": 854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960588.480, "dur": 71.626,
    "args": {
      "External id": 855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960588.818, "dur": 62.961,
    "args": {
      "External id": 856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960709.448, "dur": 18.483,
    "args": {
      "External id": 857,"Record function id": 0, "Ev Idx": 344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960718.116, "dur": 8.898,
    "args": {
      "External id": 858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960723.359, "dur": 2.668,
    "args": {
      "External id": 859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960724.638, "dur": 1.220,
    "args": {
      "External id": 860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960732.154, "dur": 5.601,
    "args": {
      "External id": 861,"Record function id": 0, "Ev Idx": 348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960733.985, "dur": 3.015,
    "args": {
      "External id": 862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960734.828, "dur": 1.632,
    "args": {
      "External id": 863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960735.862, "dur": 0.431,
    "args": {
      "External id": 864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960741.659, "dur": 4.733,
    "args": {
      "External id": 865,"Record function id": 0, "Ev Idx": 352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960743.356, "dur": 2.264,
    "args": {
      "External id": 866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960744.287, "dur": 0.942,
    "args": {
      "External id": 867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960744.752, "dur": 0.333,
    "args": {
      "External id": 868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960749.921, "dur": 4.250,
    "args": {
      "External id": 869,"Record function id": 0, "Ev Idx": 356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960751.434, "dur": 2.012,
    "args": {
      "External id": 870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960752.125, "dur": 0.975,
    "args": {
      "External id": 871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960752.545, "dur": 0.422,
    "args": {
      "External id": 872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960757.337, "dur": 3.886,
    "args": {
      "External id": 873,"Record function id": 0, "Ev Idx": 360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960758.634, "dur": 1.904,
    "args": {
      "External id": 874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960759.283, "dur": 0.910,
    "args": {
      "External id": 875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960759.691, "dur": 0.362,
    "args": {
      "External id": 876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960764.467, "dur": 4.262,
    "args": {
      "External id": 877,"Record function id": 0, "Ev Idx": 364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960765.904, "dur": 2.093,
    "args": {
      "External id": 878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960766.593, "dur": 1.032,
    "args": {
      "External id": 879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960767.052, "dur": 0.438,
    "args": {
      "External id": 880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960771.904, "dur": 4.047,
    "args": {
      "External id": 881,"Record function id": 0, "Ev Idx": 368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960773.400, "dur": 1.834,
    "args": {
      "External id": 882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960774.014, "dur": 0.852,
    "args": {
      "External id": 883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960774.401, "dur": 0.318,
    "args": {
      "External id": 884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960779.257, "dur": 4.266,
    "args": {
      "External id": 885,"Record function id": 0, "Ev Idx": 372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960780.850, "dur": 1.890,
    "args": {
      "External id": 886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960781.411, "dur": 0.946,
    "args": {
      "External id": 887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960781.847, "dur": 0.364,
    "args": {
      "External id": 888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960786.942, "dur": 4.167,
    "args": {
      "External id": 889,"Record function id": 0, "Ev Idx": 376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960788.508, "dur": 1.851,
    "args": {
      "External id": 890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960789.139, "dur": 0.831,
    "args": {
      "External id": 891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960789.511, "dur": 0.319,
    "args": {
      "External id": 892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960794.385, "dur": 4.179,
    "args": {
      "External id": 893,"Record function id": 0, "Ev Idx": 380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960795.830, "dur": 1.979,
    "args": {
      "External id": 894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960796.547, "dur": 0.882,
    "args": {
      "External id": 895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960796.927, "dur": 0.358,
    "args": {
      "External id": 896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960801.916, "dur": 4.642,
    "args": {
      "External id": 897,"Record function id": 0, "Ev Idx": 384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960803.446, "dur": 2.325,
    "args": {
      "External id": 898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960804.344, "dur": 1.038,
    "args": {
      "External id": 899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960804.887, "dur": 0.349,
    "args": {
      "External id": 900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960809.852, "dur": 3.918,
    "args": {
      "External id": 901,"Record function id": 0, "Ev Idx": 388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960811.070, "dur": 1.986,
    "args": {
      "External id": 902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960811.743, "dur": 0.920,
    "args": {
      "External id": 903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960812.110, "dur": 0.408,
    "args": {
      "External id": 904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960817.117, "dur": 3.975,
    "args": {
      "External id": 905,"Record function id": 0, "Ev Idx": 392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960818.432, "dur": 1.918,
    "args": {
      "External id": 906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960819.097, "dur": 0.863,
    "args": {
      "External id": 907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960819.468, "dur": 0.356,
    "args": {
      "External id": 908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960824.449, "dur": 5.385,
    "args": {
      "External id": 909,"Record function id": 0, "Ev Idx": 396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960825.933, "dur": 3.195,
    "args": {
      "External id": 910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960826.589, "dur": 2.119,
    "args": {
      "External id": 911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960826.952, "dur": 1.609,
    "args": {
      "External id": 912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960833.364, "dur": 4.151,
    "args": {
      "External id": 913,"Record function id": 0, "Ev Idx": 400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960834.913, "dur": 1.838,
    "args": {
      "External id": 914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960835.488, "dur": 0.883,
    "args": {
      "External id": 915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960835.851, "dur": 0.375,
    "args": {
      "External id": 916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960840.976, "dur": 4.326,
    "args": {
      "External id": 917,"Record function id": 0, "Ev Idx": 404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960842.539, "dur": 1.989,
    "args": {
      "External id": 918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960843.254, "dur": 0.914,
    "args": {
      "External id": 919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960843.634, "dur": 0.401,
    "args": {
      "External id": 920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960848.592, "dur": 4.235,
    "args": {
      "External id": 921,"Record function id": 0, "Ev Idx": 408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960850.205, "dur": 1.850,
    "args": {
      "External id": 922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960850.867, "dur": 0.825,
    "args": {
      "External id": 923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960851.247, "dur": 0.314,
    "args": {
      "External id": 924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960856.175, "dur": 4.796,
    "args": {
      "External id": 925,"Record function id": 0, "Ev Idx": 412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960857.735, "dur": 2.494,
    "args": {
      "External id": 926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960858.417, "dur": 1.438,
    "args": {
      "External id": 927,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960858.793, "dur": 0.921,
    "args": {
      "External id": 928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960864.375, "dur": 5.154,
    "args": {
      "External id": 929,"Record function id": 0, "Ev Idx": 416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960866.047, "dur": 2.693,
    "args": {
      "External id": 930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960866.840, "dur": 1.378,
    "args": {
      "External id": 931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960867.235, "dur": 0.842,
    "args": {
      "External id": 932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960872.852, "dur": 4.723,
    "args": {
      "External id": 933,"Record function id": 0, "Ev Idx": 420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960874.284, "dur": 2.514,
    "args": {
      "External id": 934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960875.079, "dur": 1.171,
    "args": {
      "External id": 935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960875.538, "dur": 0.569,
    "args": {
      "External id": 936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960881.008, "dur": 37.208,
    "args": {
      "External id": 937,"Record function id": 0, "Ev Idx": 424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960882.711, "dur": 34.079,
    "args": {
      "External id": 938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960883.319, "dur": 32.749,
    "args": {
      "External id": 939,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960883.699, "dur": 31.072,
    "args": {
      "External id": 940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960925.613, "dur": 6.551,
    "args": {
      "External id": 941,"Record function id": 0, "Ev Idx": 428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960929.061, "dur": 2.319,
    "args": {
      "External id": 942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960929.953, "dur": 1.004,
    "args": {
      "External id": 943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960930.407, "dur": 0.398,
    "args": {
      "External id": 944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960935.599, "dur": 4.292,
    "args": {
      "External id": 945,"Record function id": 0, "Ev Idx": 432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960937.286, "dur": 1.873,
    "args": {
      "External id": 946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960937.946, "dur": 0.897,
    "args": {
      "External id": 947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960938.364, "dur": 0.333,
    "args": {
      "External id": 948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960943.216, "dur": 4.561,
    "args": {
      "External id": 949,"Record function id": 0, "Ev Idx": 436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960944.956, "dur": 2.076,
    "args": {
      "External id": 950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960945.634, "dur": 0.989,
    "args": {
      "External id": 951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960946.124, "dur": 0.357,
    "args": {
      "External id": 952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960951.070, "dur": 39.639,
    "args": {
      "External id": 953,"Record function id": 0, "Ev Idx": 440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960952.562, "dur": 36.745,
    "args": {
      "External id": 954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960953.192, "dur": 35.387,
    "args": {
      "External id": 955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507960986.709, "dur": 1.623,
    "args": {
      "External id": 956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507960996.630, "dur": 6.491,
    "args": {
      "External id": 957,"Record function id": 0, "Ev Idx": 444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961000.115, "dur": 2.302,
    "args": {
      "External id": 958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961001.047, "dur": 0.959,
    "args": {
      "External id": 959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961001.466, "dur": 0.390,
    "args": {
      "External id": 960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961006.543, "dur": 4.207,
    "args": {
      "External id": 961,"Record function id": 0, "Ev Idx": 448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961007.979, "dur": 2.026,
    "args": {
      "External id": 962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961008.698, "dur": 0.919,
    "args": {
      "External id": 963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961009.125, "dur": 0.360,
    "args": {
      "External id": 964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961014.097, "dur": 4.310,
    "args": {
      "External id": 965,"Record function id": 0, "Ev Idx": 452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961015.657, "dur": 2.012,
    "args": {
      "External id": 966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961016.299, "dur": 0.972,
    "args": {
      "External id": 967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961016.751, "dur": 0.350,
    "args": {
      "External id": 968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961021.774, "dur": 4.106,
    "args": {
      "External id": 969,"Record function id": 0, "Ev Idx": 456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961023.154, "dur": 1.956,
    "args": {
      "External id": 970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961023.691, "dur": 1.009,
    "args": {
      "External id": 971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961024.141, "dur": 0.431,
    "args": {
      "External id": 972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961029.166, "dur": 4.435,
    "args": {
      "External id": 973,"Record function id": 0, "Ev Idx": 460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961030.585, "dur": 2.239,
    "args": {
      "External id": 974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961031.281, "dur": 0.991,
    "args": {
      "External id": 975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961031.747, "dur": 0.339,
    "args": {
      "External id": 976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961036.902, "dur": 4.254,
    "args": {
      "External id": 977,"Record function id": 0, "Ev Idx": 464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961038.357, "dur": 2.075,
    "args": {
      "External id": 978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961039.179, "dur": 0.842,
    "args": {
      "External id": 979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961039.592, "dur": 0.304,
    "args": {
      "External id": 980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961044.432, "dur": 4.085,
    "args": {
      "External id": 981,"Record function id": 0, "Ev Idx": 468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961045.916, "dur": 1.875,
    "args": {
      "External id": 982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961046.550, "dur": 0.857,
    "args": {
      "External id": 983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961046.952, "dur": 0.323,
    "args": {
      "External id": 984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961051.796, "dur": 4.384,
    "args": {
      "External id": 985,"Record function id": 0, "Ev Idx": 472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961053.269, "dur": 2.100,
    "args": {
      "External id": 986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961054.111, "dur": 0.878,
    "args": {
      "External id": 987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961054.475, "dur": 0.387,
    "args": {
      "External id": 988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961059.380, "dur": 3.875,
    "args": {
      "External id": 989,"Record function id": 0, "Ev Idx": 476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961060.621, "dur": 1.961,
    "args": {
      "External id": 990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961061.323, "dur": 0.876,
    "args": {
      "External id": 991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961061.690, "dur": 0.358,
    "args": {
      "External id": 992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961066.582, "dur": 3.952,
    "args": {
      "External id": 993,"Record function id": 0, "Ev Idx": 480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961067.925, "dur": 1.837,
    "args": {
      "External id": 994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961068.550, "dur": 0.905,
    "args": {
      "External id": 995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961068.927, "dur": 0.400,
    "args": {
      "External id": 996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961074.418, "dur": 4.135,
    "args": {
      "External id": 997,"Record function id": 0, "Ev Idx": 484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961075.949, "dur": 1.846,
    "args": {
      "External id": 998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961076.573, "dur": 0.840,
    "args": {
      "External id": 999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961076.943, "dur": 0.336,
    "args": {
      "External id": 1000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961086.047, "dur": 4.045,
    "args": {
      "External id": 1001,"Record function id": 0, "Ev Idx": 488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961087.355, "dur": 1.968,
    "args": {
      "External id": 1002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961088.001, "dur": 0.950,
    "args": {
      "External id": 1003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961088.382, "dur": 0.439,
    "args": {
      "External id": 1004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961093.365, "dur": 3.935,
    "args": {
      "External id": 1005,"Record function id": 0, "Ev Idx": 492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961094.609, "dur": 1.940,
    "args": {
      "External id": 1006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961095.239, "dur": 0.941,
    "args": {
      "External id": 1007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961095.644, "dur": 0.413,
    "args": {
      "External id": 1008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961100.459, "dur": 4.174,
    "args": {
      "External id": 1009,"Record function id": 0, "Ev Idx": 496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961102.075, "dur": 1.867,
    "args": {
      "External id": 1010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961102.624, "dur": 0.939,
    "args": {
      "External id": 1011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961103.001, "dur": 0.394,
    "args": {
      "External id": 1012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961108.229, "dur": 4.049,
    "args": {
      "External id": 1013,"Record function id": 0, "Ev Idx": 500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961109.804, "dur": 1.746,
    "args": {
      "External id": 1014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961110.334, "dur": 0.813,
    "args": {
      "External id": 1015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961110.696, "dur": 0.306,
    "args": {
      "External id": 1016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961115.548, "dur": 40.197,
    "args": {
      "External id": 1017,"Record function id": 0, "Ev Idx": 504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961117.015, "dur": 37.374,
    "args": {
      "External id": 1018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961117.785, "dur": 35.680,
    "args": {
      "External id": 1019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961118.175, "dur": 34.052,
    "args": {
      "External id": 1020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961163.167, "dur": 4.951,
    "args": {
      "External id": 1021,"Record function id": 0, "Ev Idx": 508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961164.914, "dur": 2.407,
    "args": {
      "External id": 1022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961165.871, "dur": 0.920,
    "args": {
      "External id": 1023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961166.301, "dur": 0.338,
    "args": {
      "External id": 1024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961349.581, "dur": 12.442,
    "args": {
      "External id": 1025,"Record function id": 0, "Ev Idx": 512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961357.642, "dur": 3.543,
    "args": {
      "External id": 1026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961358.613, "dur": 1.996,
    "args": {
      "External id": 1027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961359.115, "dur": 1.250,
    "args": {
      "External id": 1028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961366.070, "dur": 4.524,
    "args": {
      "External id": 1029,"Record function id": 0, "Ev Idx": 516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961367.974, "dur": 1.931,
    "args": {
      "External id": 1030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961368.609, "dur": 0.882,
    "args": {
      "External id": 1031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961369.018, "dur": 0.337,
    "args": {
      "External id": 1032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961373.895, "dur": 4.224,
    "args": {
      "External id": 1033,"Record function id": 0, "Ev Idx": 520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961375.560, "dur": 1.801,
    "args": {
      "External id": 1034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961376.105, "dur": 0.842,
    "args": {
      "External id": 1035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961376.496, "dur": 0.321,
    "args": {
      "External id": 1036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961381.256, "dur": 4.006,
    "args": {
      "External id": 1037,"Record function id": 0, "Ev Idx": 524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961382.803, "dur": 1.783,
    "args": {
      "External id": 1038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961383.315, "dur": 0.874,
    "args": {
      "External id": 1039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961383.736, "dur": 0.323,
    "args": {
      "External id": 1040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961388.642, "dur": 3.842,
    "args": {
      "External id": 1041,"Record function id": 0, "Ev Idx": 528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961389.975, "dur": 1.830,
    "args": {
      "External id": 1042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961390.577, "dur": 0.840,
    "args": {
      "External id": 1043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961390.951, "dur": 0.337,
    "args": {
      "External id": 1044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961395.591, "dur": 4.269,
    "args": {
      "External id": 1045,"Record function id": 0, "Ev Idx": 532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961397.170, "dur": 1.948,
    "args": {
      "External id": 1046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961397.888, "dur": 0.844,
    "args": {
      "External id": 1047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961398.285, "dur": 0.314,
    "args": {
      "External id": 1048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961404.691, "dur": 4.512,
    "args": {
      "External id": 1049,"Record function id": 0, "Ev Idx": 536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961406.302, "dur": 2.116,
    "args": {
      "External id": 1050,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961407.158, "dur": 0.941,
    "args": {
      "External id": 1051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961407.558, "dur": 0.401,
    "args": {
      "External id": 1052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961412.508, "dur": 6.389,
    "args": {
      "External id": 1053,"Record function id": 0, "Ev Idx": 540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961416.081, "dur": 2.074,
    "args": {
      "External id": 1054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961416.803, "dur": 0.961,
    "args": {
      "External id": 1055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961417.218, "dur": 0.418,
    "args": {
      "External id": 1056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961422.251, "dur": 4.173,
    "args": {
      "External id": 1057,"Record function id": 0, "Ev Idx": 544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961423.868, "dur": 1.838,
    "args": {
      "External id": 1058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961424.389, "dur": 0.888,
    "args": {
      "External id": 1059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961424.770, "dur": 0.359,
    "args": {
      "External id": 1060,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961429.869, "dur": 4.221,
    "args": {
      "External id": 1061,"Record function id": 0, "Ev Idx": 548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961431.339, "dur": 2.079,
    "args": {
      "External id": 1062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961432.129, "dur": 0.905,
    "args": {
      "External id": 1063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961432.533, "dur": 0.352,
    "args": {
      "External id": 1064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961437.413, "dur": 4.470,
    "args": {
      "External id": 1065,"Record function id": 0, "Ev Idx": 552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961438.972, "dur": 2.166,
    "args": {
      "External id": 1066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961439.739, "dur": 0.823,
    "args": {
      "External id": 1067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961440.110, "dur": 0.318,
    "args": {
      "External id": 1068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961445.175, "dur": 3.798,
    "args": {
      "External id": 1069,"Record function id": 0, "Ev Idx": 556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961446.608, "dur": 1.669,
    "args": {
      "External id": 1070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961447.134, "dur": 0.828,
    "args": {
      "External id": 1071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961447.505, "dur": 0.329,
    "args": {
      "External id": 1072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961452.343, "dur": 4.726,
    "args": {
      "External id": 1073,"Record function id": 0, "Ev Idx": 560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961453.872, "dur": 2.477,
    "args": {
      "External id": 1074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961454.404, "dur": 1.545,
    "args": {
      "External id": 1075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961454.819, "dur": 0.996,
    "args": {
      "External id": 1076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961460.278, "dur": 4.020,
    "args": {
      "External id": 1077,"Record function id": 0, "Ev Idx": 564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961461.770, "dur": 1.868,
    "args": {
      "External id": 1078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961462.359, "dur": 0.885,
    "args": {
      "External id": 1079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961462.739, "dur": 0.378,
    "args": {
      "External id": 1080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961467.654, "dur": 3.891,
    "args": {
      "External id": 1081,"Record function id": 0, "Ev Idx": 568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961468.978, "dur": 1.813,
    "args": {
      "External id": 1082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961469.503, "dur": 0.893,
    "args": {
      "External id": 1083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961469.902, "dur": 0.365,
    "args": {
      "External id": 1084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961474.715, "dur": 4.072,
    "args": {
      "External id": 1085,"Record function id": 0, "Ev Idx": 572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961476.131, "dur": 1.855,
    "args": {
      "External id": 1086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961476.725, "dur": 0.874,
    "args": {
      "External id": 1087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961477.096, "dur": 0.376,
    "args": {
      "External id": 1088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961482.034, "dur": 4.260,
    "args": {
      "External id": 1089,"Record function id": 0, "Ev Idx": 576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961483.415, "dur": 2.113,
    "args": {
      "External id": 1090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961483.969, "dur": 1.167,
    "args": {
      "External id": 1091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961484.342, "dur": 0.666,
    "args": {
      "External id": 1092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961489.594, "dur": 40.697,
    "args": {
      "External id": 1093,"Record function id": 0, "Ev Idx": 580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961525.093, "dur": 3.879,
    "args": {
      "External id": 1094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961526.418, "dur": 1.955,
    "args": {
      "External id": 1095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961526.873, "dur": 1.289,
    "args": {
      "External id": 1096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961534.447, "dur": 4.038,
    "args": {
      "External id": 1097,"Record function id": 0, "Ev Idx": 584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961536.039, "dur": 1.699,
    "args": {
      "External id": 1098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961536.601, "dur": 0.825,
    "args": {
      "External id": 1099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961536.966, "dur": 0.326,
    "args": {
      "External id": 1100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961542.939, "dur": 46.367,
    "args": {
      "External id": 1101,"Record function id": 0, "Ev Idx": 588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961544.606, "dur": 42.654,
    "args": {
      "External id": 1102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961546.124, "dur": 40.207,
    "args": {
      "External id": 1103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961546.517, "dur": 38.623,
    "args": {
      "External id": 1104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961595.489, "dur": 7.165,
    "args": {
      "External id": 1105,"Record function id": 0, "Ev Idx": 592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961599.308, "dur": 2.550,
    "args": {
      "External id": 1106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961600.251, "dur": 1.081,
    "args": {
      "External id": 1107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961600.689, "dur": 0.479,
    "args": {
      "External id": 1108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961606.001, "dur": 4.083,
    "args": {
      "External id": 1109,"Record function id": 0, "Ev Idx": 596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961607.509, "dur": 1.866,
    "args": {
      "External id": 1110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961608.094, "dur": 0.879,
    "args": {
      "External id": 1111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961608.470, "dur": 0.349,
    "args": {
      "External id": 1112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961613.480, "dur": 4.229,
    "args": {
      "External id": 1113,"Record function id": 0, "Ev Idx": 600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961615.234, "dur": 1.780,
    "args": {
      "External id": 1114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961615.783, "dur": 0.843,
    "args": {
      "External id": 1115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961616.124, "dur": 0.357,
    "args": {
      "External id": 1116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961621.075, "dur": 4.252,
    "args": {
      "External id": 1117,"Record function id": 0, "Ev Idx": 604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961622.713, "dur": 1.866,
    "args": {
      "External id": 1118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961623.336, "dur": 0.853,
    "args": {
      "External id": 1119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961623.686, "dur": 0.327,
    "args": {
      "External id": 1120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961628.551, "dur": 4.959,
    "args": {
      "External id": 1121,"Record function id": 0, "Ev Idx": 608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961630.938, "dur": 1.820,
    "args": {
      "External id": 1122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961631.523, "dur": 0.844,
    "args": {
      "External id": 1123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961631.873, "dur": 0.340,
    "args": {
      "External id": 1124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961636.760, "dur": 4.230,
    "args": {
      "External id": 1125,"Record function id": 0, "Ev Idx": 612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961638.150, "dur": 2.107,
    "args": {
      "External id": 1126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961638.727, "dur": 1.138,
    "args": {
      "External id": 1127,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961639.229, "dur": 0.491,
    "args": {
      "External id": 1128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961644.284, "dur": 5.046,
    "args": {
      "External id": 1129,"Record function id": 0, "Ev Idx": 616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961646.620, "dur": 2.006,
    "args": {
      "External id": 1130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961647.260, "dur": 0.949,
    "args": {
      "External id": 1131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961647.701, "dur": 0.373,
    "args": {
      "External id": 1132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961652.620, "dur": 4.394,
    "args": {
      "External id": 1133,"Record function id": 0, "Ev Idx": 620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961654.276, "dur": 2.040,
    "args": {
      "External id": 1134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961654.901, "dur": 1.025,
    "args": {
      "External id": 1135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961655.298, "dur": 0.485,
    "args": {
      "External id": 1136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961660.316, "dur": 4.186,
    "args": {
      "External id": 1137,"Record function id": 0, "Ev Idx": 624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961661.899, "dur": 1.821,
    "args": {
      "External id": 1138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961662.472, "dur": 0.854,
    "args": {
      "External id": 1139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961662.834, "dur": 0.347,
    "args": {
      "External id": 1140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961667.982, "dur": 3.919,
    "args": {
      "External id": 1141,"Record function id": 0, "Ev Idx": 628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961669.375, "dur": 1.750,
    "args": {
      "External id": 1142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961669.921, "dur": 0.812,
    "args": {
      "External id": 1143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961670.252, "dur": 0.352,
    "args": {
      "External id": 1144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961675.199, "dur": 4.700,
    "args": {
      "External id": 1145,"Record function id": 0, "Ev Idx": 632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961677.029, "dur": 2.070,
    "args": {
      "External id": 1146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961677.726, "dur": 0.868,
    "args": {
      "External id": 1147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961678.106, "dur": 0.345,
    "args": {
      "External id": 1148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961683.165, "dur": 4.236,
    "args": {
      "External id": 1149,"Record function id": 0, "Ev Idx": 636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961684.596, "dur": 2.037,
    "args": {
      "External id": 1150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961685.276, "dur": 0.852,
    "args": {
      "External id": 1151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961685.640, "dur": 0.348,
    "args": {
      "External id": 1152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961690.510, "dur": 4.357,
    "args": {
      "External id": 1153,"Record function id": 0, "Ev Idx": 640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961692.271, "dur": 1.811,
    "args": {
      "External id": 1154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961692.861, "dur": 0.822,
    "args": {
      "External id": 1155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961693.224, "dur": 0.321,
    "args": {
      "External id": 1156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961700.021, "dur": 4.312,
    "args": {
      "External id": 1157,"Record function id": 0, "Ev Idx": 644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961701.637, "dur": 1.947,
    "args": {
      "External id": 1158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961702.268, "dur": 0.920,
    "args": {
      "External id": 1159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961702.643, "dur": 0.418,
    "args": {
      "External id": 1160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961707.393, "dur": 4.658,
    "args": {
      "External id": 1161,"Record function id": 0, "Ev Idx": 648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961708.918, "dur": 2.453,
    "args": {
      "External id": 1162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961709.528, "dur": 1.450,
    "args": {
      "External id": 1163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961709.872, "dur": 0.964,
    "args": {
      "External id": 1164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961715.143, "dur": 4.152,
    "args": {
      "External id": 1165,"Record function id": 0, "Ev Idx": 652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961716.766, "dur": 1.836,
    "args": {
      "External id": 1166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961717.385, "dur": 0.823,
    "args": {
      "External id": 1167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961717.732, "dur": 0.352,
    "args": {
      "External id": 1168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961722.657, "dur": 3.785,
    "args": {
      "External id": 1169,"Record function id": 0, "Ev Idx": 656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961723.956, "dur": 1.794,
    "args": {
      "External id": 1170,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961724.520, "dur": 0.807,
    "args": {
      "External id": 1171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961724.864, "dur": 0.336,
    "args": {
      "External id": 1172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961729.539, "dur": 4.514,
    "args": {
      "External id": 1173,"Record function id": 0, "Ev Idx": 660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961731.016, "dur": 2.358,
    "args": {
      "External id": 1174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961731.680, "dur": 1.309,
    "args": {
      "External id": 1175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961732.019, "dur": 0.845,
    "args": {
      "External id": 1176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961737.270, "dur": 4.113,
    "args": {
      "External id": 1177,"Record function id": 0, "Ev Idx": 664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961738.583, "dur": 2.120,
    "args": {
      "External id": 1178,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961739.149, "dur": 1.149,
    "args": {
      "External id": 1179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961739.505, "dur": 0.667,
    "args": {
      "External id": 1180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961744.745, "dur": 4.447,
    "args": {
      "External id": 1181,"Record function id": 0, "Ev Idx": 668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961746.416, "dur": 1.986,
    "args": {
      "External id": 1182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961747.033, "dur": 0.962,
    "args": {
      "External id": 1183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961747.379, "dur": 0.488,
    "args": {
      "External id": 1184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961752.581, "dur": 37.305,
    "args": {
      "External id": 1185,"Record function id": 0, "Ev Idx": 672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961754.298, "dur": 34.256,
    "args": {
      "External id": 1186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961755.131, "dur": 32.421,
    "args": {
      "External id": 1187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961755.475, "dur": 30.850,
    "args": {
      "External id": 1188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961796.146, "dur": 7.458,
    "args": {
      "External id": 1189,"Record function id": 0, "Ev Idx": 676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961800.364, "dur": 2.446,
    "args": {
      "External id": 1190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961801.370, "dur": 0.937,
    "args": {
      "External id": 1191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961801.796, "dur": 0.363,
    "args": {
      "External id": 1192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961806.945, "dur": 4.266,
    "args": {
      "External id": 1193,"Record function id": 0, "Ev Idx": 680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961808.638, "dur": 1.850,
    "args": {
      "External id": 1194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961809.226, "dur": 0.861,
    "args": {
      "External id": 1195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961809.628, "dur": 0.326,
    "args": {
      "External id": 1196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961814.431, "dur": 3.725,
    "args": {
      "External id": 1197,"Record function id": 0, "Ev Idx": 684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961815.766, "dur": 1.722,
    "args": {
      "External id": 1198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961816.338, "dur": 0.817,
    "args": {
      "External id": 1199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961816.701, "dur": 0.322,
    "args": {
      "External id": 1200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961821.351, "dur": 3.997,
    "args": {
      "External id": 1201,"Record function id": 0, "Ev Idx": 688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961822.885, "dur": 1.777,
    "args": {
      "External id": 1202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961823.470, "dur": 0.839,
    "args": {
      "External id": 1203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961823.824, "dur": 0.338,
    "args": {
      "External id": 1204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961828.607, "dur": 3.845,
    "args": {
      "External id": 1205,"Record function id": 0, "Ev Idx": 692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961829.972, "dur": 1.755,
    "args": {
      "External id": 1206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961830.526, "dur": 0.803,
    "args": {
      "External id": 1207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961830.877, "dur": 0.325,
    "args": {
      "External id": 1208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961835.832, "dur": 4.861,
    "args": {
      "External id": 1209,"Record function id": 0, "Ev Idx": 696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961837.427, "dur": 2.540,
    "args": {
      "External id": 1210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961838.725, "dur": 0.818,
    "args": {
      "External id": 1211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961839.069, "dur": 0.344,
    "args": {
      "External id": 1212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961843.840, "dur": 3.926,
    "args": {
      "External id": 1213,"Record function id": 0, "Ev Idx": 700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961845.220, "dur": 1.816,
    "args": {
      "External id": 1214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961845.787, "dur": 0.867,
    "args": {
      "External id": 1215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961846.131, "dur": 0.375,
    "args": {
      "External id": 1216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961851.009, "dur": 3.827,
    "args": {
      "External id": 1217,"Record function id": 0, "Ev Idx": 704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961852.354, "dur": 1.759,
    "args": {
      "External id": 1218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961852.905, "dur": 0.818,
    "args": {
      "External id": 1219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961853.275, "dur": 0.320,
    "args": {
      "External id": 1220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961857.876, "dur": 3.826,
    "args": {
      "External id": 1221,"Record function id": 0, "Ev Idx": 708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961859.233, "dur": 1.787,
    "args": {
      "External id": 1222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961859.799, "dur": 0.832,
    "args": {
      "External id": 1223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961860.139, "dur": 0.363,
    "args": {
      "External id": 1224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961864.853, "dur": 3.966,
    "args": {
      "External id": 1225,"Record function id": 0, "Ev Idx": 712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961866.271, "dur": 1.871,
    "args": {
      "External id": 1226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961866.916, "dur": 0.831,
    "args": {
      "External id": 1227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961867.275, "dur": 0.339,
    "args": {
      "External id": 1228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961871.915, "dur": 4.057,
    "args": {
      "External id": 1229,"Record function id": 0, "Ev Idx": 716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961873.489, "dur": 1.813,
    "args": {
      "External id": 1230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961874.053, "dur": 0.858,
    "args": {
      "External id": 1231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961874.412, "dur": 0.359,
    "args": {
      "External id": 1232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961879.331, "dur": 3.995,
    "args": {
      "External id": 1233,"Record function id": 0, "Ev Idx": 720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961880.867, "dur": 1.788,
    "args": {
      "External id": 1234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961881.435, "dur": 0.826,
    "args": {
      "External id": 1235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961881.770, "dur": 0.364,
    "args": {
      "External id": 1236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961886.298, "dur": 5.608,
    "args": {
      "External id": 1237,"Record function id": 0, "Ev Idx": 724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961888.808, "dur": 2.420,
    "args": {
      "External id": 1238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961889.432, "dur": 1.385,
    "args": {
      "External id": 1239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961889.804, "dur": 0.883,
    "args": {
      "External id": 1240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961895.148, "dur": 4.466,
    "args": {
      "External id": 1241,"Record function id": 0, "Ev Idx": 728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961896.890, "dur": 1.963,
    "args": {
      "External id": 1242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961897.483, "dur": 0.960,
    "args": {
      "External id": 1243,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961897.870, "dur": 0.381,
    "args": {
      "External id": 1244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961902.846, "dur": 4.379,
    "args": {
      "External id": 1245,"Record function id": 0, "Ev Idx": 732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961904.519, "dur": 1.974,
    "args": {
      "External id": 1246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961905.215, "dur": 0.887,
    "args": {
      "External id": 1247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961905.557, "dur": 0.408,
    "args": {
      "External id": 1248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961910.290, "dur": 4.009,
    "args": {
      "External id": 1249,"Record function id": 0, "Ev Idx": 736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961911.678, "dur": 1.868,
    "args": {
      "External id": 1250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961912.244, "dur": 0.912,
    "args": {
      "External id": 1251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961912.591, "dur": 0.442,
    "args": {
      "External id": 1252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961917.432, "dur": 37.163,
    "args": {
      "External id": 1253,"Record function id": 0, "Ev Idx": 740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961919.162, "dur": 34.074,
    "args": {
      "External id": 1254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961919.907, "dur": 31.712,
    "args": {
      "External id": 1255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961920.283, "dur": 30.097,
    "args": {
      "External id": 1256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961960.517, "dur": 7.729,
    "args": {
      "External id": 1257,"Record function id": 0, "Ev Idx": 744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961964.874, "dur": 2.622,
    "args": {
      "External id": 1258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961965.933, "dur": 1.116,
    "args": {
      "External id": 1259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961966.486, "dur": 0.434,
    "args": {
      "External id": 1260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961971.419, "dur": 4.415,
    "args": {
      "External id": 1261,"Record function id": 0, "Ev Idx": 748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961973.132, "dur": 1.954,
    "args": {
      "External id": 1262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961973.774, "dur": 0.935,
    "args": {
      "External id": 1263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961974.227, "dur": 0.359,
    "args": {
      "External id": 1264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961979.865, "dur": 4.218,
    "args": {
      "External id": 1265,"Record function id": 0, "Ev Idx": 752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961981.452, "dur": 1.884,
    "args": {
      "External id": 1266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961982.018, "dur": 0.906,
    "args": {
      "External id": 1267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961982.455, "dur": 0.318,
    "args": {
      "External id": 1268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961987.061, "dur": 4.054,
    "args": {
      "External id": 1269,"Record function id": 0, "Ev Idx": 756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961988.626, "dur": 1.799,
    "args": {
      "External id": 1270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961989.195, "dur": 0.830,
    "args": {
      "External id": 1271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961989.547, "dur": 0.348,
    "args": {
      "External id": 1272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961994.411, "dur": 4.238,
    "args": {
      "External id": 1273,"Record function id": 0, "Ev Idx": 760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507961996.147, "dur": 1.807,
    "args": {
      "External id": 1274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961996.719, "dur": 0.840,
    "args": {
      "External id": 1275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507961997.055, "dur": 0.372,
    "args": {
      "External id": 1276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507962001.599, "dur": 4.047,
    "args": {
      "External id": 1277,"Record function id": 0, "Ev Idx": 764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507962002.993, "dur": 1.960,
    "args": {
      "External id": 1278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507962003.645, "dur": 0.844,
    "args": {
      "External id": 1279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507962004.019, "dur": 0.344,
    "args": {
      "External id": 1280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507962009.147, "dur": 253.041,
    "args": {
      "External id": 1281,"Record function id": 0, "Sequence number": 4684, "Fwd thread id": 1, "Ev Idx": 768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507962010.874, "dur": 155.852,
    "args": {
      "External id": 1282,"Sequence number": 4684, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 769
    }
  },
  {
    "ph": "f", "id": 2, "pid": 14565, "tid": 14640, "ts": 1290507962010.874,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290507962084.586, "dur": 68.697,
    "args": {
      "External id": 1283,"Record function id": 0, "scope": "user_scope", "Ev Idx": 770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290507962181.676, "dur": 76.304,
    "args": {
      "External id": 1284,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507962271.337, "dur": 9.695,
    "args": {
      "External id": 1285,"Record function id": 0, "Ev Idx": 772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507962276.506, "dur": 3.531,
    "args": {
      "External id": 1286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507962277.857, "dur": 1.833,
    "args": {
      "External id": 1287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507962278.621, "dur": 0.900,
    "args": {
      "External id": 1288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507964872.873, "dur": 2094.401,
    "args": {
      "External id": 1289,"Record function id": 0, "Sequence number": 4687, "Fwd thread id": 1, "Ev Idx": 776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507964875.702, "dur": 1994.704,
    "args": {
      "External id": 1290,"Sequence number": 4687, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 777
    }
  },
  {
    "ph": "f", "id": 3, "pid": 14565, "tid": 14640, "ts": 1290507964875.702,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290507965202.546, "dur": 1640.547,
    "args": {
      "External id": 1291,"Record function id": 0, "scope": "user_scope", "Ev Idx": 778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290507965879.089, "dur": 25.761,
    "args": {
      "External id": 1292,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507966992.022, "dur": 14.357,
    "args": {
      "External id": 1293,"Record function id": 0, "Ev Idx": 780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507966996.571, "dur": 8.327,
    "args": {
      "External id": 1294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967000.639, "dur": 3.656,
    "args": {
      "External id": 1295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967001.951, "dur": 2.146,
    "args": {
      "External id": 1296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967010.524, "dur": 5.148,
    "args": {
      "External id": 1297,"Record function id": 0, "Ev Idx": 784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967012.068, "dur": 2.726,
    "args": {
      "External id": 1298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967013.180, "dur": 1.276,
    "args": {
      "External id": 1299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967013.755, "dur": 0.544,
    "args": {
      "External id": 1300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967019.254, "dur": 4.195,
    "args": {
      "External id": 1301,"Record function id": 0, "Ev Idx": 788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967020.704, "dur": 1.980,
    "args": {
      "External id": 1302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967021.382, "dur": 0.986,
    "args": {
      "External id": 1303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967021.844, "dur": 0.393,
    "args": {
      "External id": 1304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967026.886, "dur": 4.182,
    "args": {
      "External id": 1305,"Record function id": 0, "Ev Idx": 792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967028.184, "dur": 2.143,
    "args": {
      "External id": 1306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967028.961, "dur": 1.071,
    "args": {
      "External id": 1307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967029.528, "dur": 0.355,
    "args": {
      "External id": 1308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967034.307, "dur": 4.054,
    "args": {
      "External id": 1309,"Record function id": 0, "Ev Idx": 796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967035.623, "dur": 1.965,
    "args": {
      "External id": 1310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967036.348, "dur": 0.905,
    "args": {
      "External id": 1311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967036.724, "dur": 0.387,
    "args": {
      "External id": 1312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967041.578, "dur": 4.412,
    "args": {
      "External id": 1313,"Record function id": 0, "Ev Idx": 800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967043.091, "dur": 2.146,
    "args": {
      "External id": 1314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967043.767, "dur": 1.148,
    "args": {
      "External id": 1315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967044.280, "dur": 0.475,
    "args": {
      "External id": 1316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967049.233, "dur": 4.203,
    "args": {
      "External id": 1317,"Record function id": 0, "Ev Idx": 804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967050.514, "dur": 2.171,
    "args": {
      "External id": 1318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967051.361, "dur": 1.015,
    "args": {
      "External id": 1319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967051.833, "dur": 0.376,
    "args": {
      "External id": 1320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967056.703, "dur": 3.925,
    "args": {
      "External id": 1321,"Record function id": 0, "Ev Idx": 808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967058.073, "dur": 1.815,
    "args": {
      "External id": 1322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967058.765, "dur": 0.823,
    "args": {
      "External id": 1323,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967059.125, "dur": 0.338,
    "args": {
      "External id": 1324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967063.943, "dur": 102.411,
    "args": {
      "External id": 1325,"Record function id": 0, "Ev Idx": 812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967065.345, "dur": 95.882,
    "args": {
      "External id": 1326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967066.062, "dur": 91.915,
    "args": {
      "External id": 1327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967066.432, "dur": 87.578,
    "args": {
      "External id": 1328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967194.747, "dur": 17.462,
    "args": {
      "External id": 1329,"Record function id": 0, "Ev Idx": 816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967203.216, "dur": 8.106,
    "args": {
      "External id": 1330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967207.137, "dur": 2.861,
    "args": {
      "External id": 1331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967208.397, "dur": 1.440,
    "args": {
      "External id": 1332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967216.726, "dur": 5.430,
    "args": {
      "External id": 1333,"Record function id": 0, "Ev Idx": 820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967218.399, "dur": 2.994,
    "args": {
      "External id": 1334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967219.365, "dur": 1.429,
    "args": {
      "External id": 1335,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967220.269, "dur": 0.382,
    "args": {
      "External id": 1336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967225.763, "dur": 4.753,
    "args": {
      "External id": 1337,"Record function id": 0, "Ev Idx": 824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967227.476, "dur": 2.292,
    "args": {
      "External id": 1338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967228.536, "dur": 0.828,
    "args": {
      "External id": 1339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967228.908, "dur": 0.328,
    "args": {
      "External id": 1340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967233.948, "dur": 5.712,
    "args": {
      "External id": 1341,"Record function id": 0, "Ev Idx": 828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967236.479, "dur": 2.339,
    "args": {
      "External id": 1342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967237.333, "dur": 1.074,
    "args": {
      "External id": 1343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967237.817, "dur": 0.457,
    "args": {
      "External id": 1344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967245.238, "dur": 4.431,
    "args": {
      "External id": 1345,"Record function id": 0, "Ev Idx": 832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967246.840, "dur": 2.066,
    "args": {
      "External id": 1346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967247.448, "dur": 1.043,
    "args": {
      "External id": 1347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967247.934, "dur": 0.434,
    "args": {
      "External id": 1348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967253.021, "dur": 4.159,
    "args": {
      "External id": 1349,"Record function id": 0, "Ev Idx": 836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967254.576, "dur": 1.842,
    "args": {
      "External id": 1350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967255.218, "dur": 0.816,
    "args": {
      "External id": 1351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967255.553, "dur": 0.344,
    "args": {
      "External id": 1352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967260.519, "dur": 3.843,
    "args": {
      "External id": 1353,"Record function id": 0, "Ev Idx": 840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967261.777, "dur": 1.825,
    "args": {
      "External id": 1354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967262.388, "dur": 0.830,
    "args": {
      "External id": 1355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967262.767, "dur": 0.320,
    "args": {
      "External id": 1356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967267.656, "dur": 4.055,
    "args": {
      "External id": 1357,"Record function id": 0, "Ev Idx": 844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967268.972, "dur": 1.956,
    "args": {
      "External id": 1358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967269.669, "dur": 0.872,
    "args": {
      "External id": 1359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967270.059, "dur": 0.345,
    "args": {
      "External id": 1360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967274.955, "dur": 4.038,
    "args": {
      "External id": 1361,"Record function id": 0, "Ev Idx": 848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967276.221, "dur": 2.062,
    "args": {
      "External id": 1362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967277.031, "dur": 0.866,
    "args": {
      "External id": 1363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967277.381, "dur": 0.385,
    "args": {
      "External id": 1364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967282.222, "dur": 3.843,
    "args": {
      "External id": 1365,"Record function id": 0, "Ev Idx": 852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967283.498, "dur": 1.862,
    "args": {
      "External id": 1366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967284.112, "dur": 0.860,
    "args": {
      "External id": 1367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967284.520, "dur": 0.319,
    "args": {
      "External id": 1368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967289.310, "dur": 4.057,
    "args": {
      "External id": 1369,"Record function id": 0, "Ev Idx": 856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967290.692, "dur": 1.910,
    "args": {
      "External id": 1370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967291.409, "dur": 0.803,
    "args": {
      "External id": 1371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967291.754, "dur": 0.323,
    "args": {
      "External id": 1372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967296.559, "dur": 4.450,
    "args": {
      "External id": 1373,"Record function id": 0, "Ev Idx": 860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967298.078, "dur": 2.186,
    "args": {
      "External id": 1374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967298.949, "dur": 0.816,
    "args": {
      "External id": 1375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967299.332, "dur": 0.305,
    "args": {
      "External id": 1376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967304.243, "dur": 4.207,
    "args": {
      "External id": 1377,"Record function id": 0, "Ev Idx": 864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967305.626, "dur": 2.099,
    "args": {
      "External id": 1378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967306.366, "dur": 0.811,
    "args": {
      "External id": 1379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967306.727, "dur": 0.311,
    "args": {
      "External id": 1380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967311.720, "dur": 4.968,
    "args": {
      "External id": 1381,"Record function id": 0, "Ev Idx": 868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967313.800, "dur": 2.138,
    "args": {
      "External id": 1382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967314.509, "dur": 1.071,
    "args": {
      "External id": 1383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967314.877, "dur": 0.563,
    "args": {
      "External id": 1384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967320.010, "dur": 4.478,
    "args": {
      "External id": 1385,"Record function id": 0, "Ev Idx": 872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967321.542, "dur": 2.212,
    "args": {
      "External id": 1386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967322.226, "dur": 1.139,
    "args": {
      "External id": 1387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967322.741, "dur": 0.493,
    "args": {
      "External id": 1388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967327.869, "dur": 4.214,
    "args": {
      "External id": 1389,"Record function id": 0, "Ev Idx": 876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967329.249, "dur": 2.092,
    "args": {
      "External id": 1390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967329.951, "dur": 0.998,
    "args": {
      "External id": 1391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967330.413, "dur": 0.385,
    "args": {
      "External id": 1392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967335.462, "dur": 3.838,
    "args": {
      "External id": 1393,"Record function id": 0, "Ev Idx": 880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967336.748, "dur": 1.801,
    "args": {
      "External id": 1394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967337.328, "dur": 0.840,
    "args": {
      "External id": 1395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967337.698, "dur": 0.333,
    "args": {
      "External id": 1396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967342.535, "dur": 109.480,
    "args": {
      "External id": 1397,"Record function id": 0, "Ev Idx": 884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967425.109, "dur": 21.907,
    "args": {
      "External id": 1398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967438.389, "dur": 6.734,
    "args": {
      "External id": 1399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967440.092, "dur": 3.949,
    "args": {
      "External id": 1400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967471.179, "dur": 8.591,
    "args": {
      "External id": 1401,"Record function id": 0, "Ev Idx": 888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967475.132, "dur": 3.770,
    "args": {
      "External id": 1402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967476.189, "dur": 2.183,
    "args": {
      "External id": 1403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967477.193, "dur": 1.020,
    "args": {
      "External id": 1404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967485.014, "dur": 114.457,
    "args": {
      "External id": 1405,"Record function id": 0, "Ev Idx": 892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967486.945, "dur": 105.451,
    "args": {
      "External id": 1406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967488.011, "dur": 99.085,
    "args": {
      "External id": 1407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967488.499, "dur": 89.971,
    "args": {
      "External id": 1408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967643.397, "dur": 22.413,
    "args": {
      "External id": 1409,"Record function id": 0, "Ev Idx": 896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967654.053, "dur": 10.680,
    "args": {
      "External id": 1410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967660.113, "dur": 4.078,
    "args": {
      "External id": 1411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967662.224, "dur": 1.794,
    "args": {
      "External id": 1412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967671.100, "dur": 7.985,
    "args": {
      "External id": 1413,"Record function id": 0, "Ev Idx": 900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967674.106, "dur": 4.242,
    "args": {
      "External id": 1414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967675.477, "dur": 2.282,
    "args": {
      "External id": 1415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967677.060, "dur": 0.559,
    "args": {
      "External id": 1416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967682.901, "dur": 8.369,
    "args": {
      "External id": 1417,"Record function id": 0, "Ev Idx": 904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967685.868, "dur": 4.535,
    "args": {
      "External id": 1418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967687.020, "dur": 1.662,
    "args": {
      "External id": 1419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967688.207, "dur": 0.306,
    "args": {
      "External id": 1420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967694.817, "dur": 5.362,
    "args": {
      "External id": 1421,"Record function id": 0, "Ev Idx": 908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967696.668, "dur": 2.648,
    "args": {
      "External id": 1422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967697.827, "dur": 0.880,
    "args": {
      "External id": 1423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967698.218, "dur": 0.350,
    "args": {
      "External id": 1424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967703.687, "dur": 4.873,
    "args": {
      "External id": 1425,"Record function id": 0, "Ev Idx": 912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967705.454, "dur": 2.385,
    "args": {
      "External id": 1426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967706.401, "dur": 0.944,
    "args": {
      "External id": 1427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967706.917, "dur": 0.301,
    "args": {
      "External id": 1428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967712.217, "dur": 5.221,
    "args": {
      "External id": 1429,"Record function id": 0, "Ev Idx": 916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967714.196, "dur": 2.495,
    "args": {
      "External id": 1430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967715.121, "dur": 1.035,
    "args": {
      "External id": 1431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967715.502, "dur": 0.509,
    "args": {
      "External id": 1432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967721.060, "dur": 5.307,
    "args": {
      "External id": 1433,"Record function id": 0, "Ev Idx": 920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967723.327, "dur": 2.313,
    "args": {
      "External id": 1434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967724.138, "dur": 0.966,
    "args": {
      "External id": 1435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967724.624, "dur": 0.339,
    "args": {
      "External id": 1436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967729.865, "dur": 4.360,
    "args": {
      "External id": 1437,"Record function id": 0, "Ev Idx": 924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967731.420, "dur": 2.105,
    "args": {
      "External id": 1438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967732.174, "dur": 0.831,
    "args": {
      "External id": 1439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967732.549, "dur": 0.327,
    "args": {
      "External id": 1440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967737.537, "dur": 4.499,
    "args": {
      "External id": 1441,"Record function id": 0, "Ev Idx": 928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967739.202, "dur": 2.129,
    "args": {
      "External id": 1442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967739.962, "dur": 0.859,
    "args": {
      "External id": 1443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967740.367, "dur": 0.316,
    "args": {
      "External id": 1444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967745.248, "dur": 4.457,
    "args": {
      "External id": 1445,"Record function id": 0, "Ev Idx": 932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967746.694, "dur": 2.306,
    "args": {
      "External id": 1446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967747.479, "dur": 1.009,
    "args": {
      "External id": 1447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967748.010, "dur": 0.342,
    "args": {
      "External id": 1448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967753.055, "dur": 4.712,
    "args": {
      "External id": 1449,"Record function id": 0, "Ev Idx": 936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967754.727, "dur": 2.342,
    "args": {
      "External id": 1450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967755.495, "dur": 1.012,
    "args": {
      "External id": 1451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967756.004, "dur": 0.378,
    "args": {
      "External id": 1452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967760.964, "dur": 4.862,
    "args": {
      "External id": 1453,"Record function id": 0, "Ev Idx": 940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967762.514, "dur": 2.504,
    "args": {
      "External id": 1454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967763.444, "dur": 1.073,
    "args": {
      "External id": 1455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967763.952, "dur": 0.422,
    "args": {
      "External id": 1456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967769.218, "dur": 6.656,
    "args": {
      "External id": 1457,"Record function id": 0, "Ev Idx": 944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967770.876, "dur": 4.279,
    "args": {
      "External id": 1458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967771.634, "dur": 2.989,
    "args": {
      "External id": 1459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967772.005, "dur": 2.473,
    "args": {
      "External id": 1460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967779.308, "dur": 5.091,
    "args": {
      "External id": 1461,"Record function id": 0, "Ev Idx": 948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967781.115, "dur": 2.531,
    "args": {
      "External id": 1462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967782.027, "dur": 1.000,
    "args": {
      "External id": 1463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967782.549, "dur": 0.338,
    "args": {
      "External id": 1464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967787.918, "dur": 5.613,
    "args": {
      "External id": 1465,"Record function id": 0, "Ev Idx": 952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967790.267, "dur": 2.394,
    "args": {
      "External id": 1466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967791.139, "dur": 0.957,
    "args": {
      "External id": 1467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967791.508, "dur": 0.447,
    "args": {
      "External id": 1468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967796.845, "dur": 4.825,
    "args": {
      "External id": 1469,"Record function id": 0, "Ev Idx": 956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967798.726, "dur": 2.107,
    "args": {
      "External id": 1470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967799.495, "dur": 0.834,
    "args": {
      "External id": 1471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967799.873, "dur": 0.324,
    "args": {
      "External id": 1472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967805.030, "dur": 5.532,
    "args": {
      "External id": 1473,"Record function id": 0, "Ev Idx": 960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967806.656, "dur": 3.061,
    "args": {
      "External id": 1474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967807.538, "dur": 1.662,
    "args": {
      "External id": 1475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967807.912, "dur": 1.153,
    "args": {
      "External id": 1476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967813.897, "dur": 5.585,
    "args": {
      "External id": 1477,"Record function id": 0, "Ev Idx": 964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967815.491, "dur": 3.147,
    "args": {
      "External id": 1478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967816.341, "dur": 1.800,
    "args": {
      "External id": 1479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967816.759, "dur": 1.200,
    "args": {
      "External id": 1480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967823.054, "dur": 4.968,
    "args": {
      "External id": 1481,"Record function id": 0, "Ev Idx": 968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967824.795, "dur": 2.355,
    "args": {
      "External id": 1482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967825.555, "dur": 1.075,
    "args": {
      "External id": 1483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967826.058, "dur": 0.430,
    "args": {
      "External id": 1484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967831.510, "dur": 4.726,
    "args": {
      "External id": 1485,"Record function id": 0, "Ev Idx": 972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967832.932, "dur": 2.474,
    "args": {
      "External id": 1486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967833.647, "dur": 1.217,
    "args": {
      "External id": 1487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967834.030, "dur": 0.694,
    "args": {
      "External id": 1488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967839.912, "dur": 56.381,
    "args": {
      "External id": 1489,"Record function id": 0, "Ev Idx": 976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967841.428, "dur": 50.515,
    "args": {
      "External id": 1490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967842.197, "dur": 48.860,
    "args": {
      "External id": 1491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967842.612, "dur": 47.057,
    "args": {
      "External id": 1492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967906.483, "dur": 8.223,
    "args": {
      "External id": 1493,"Record function id": 0, "Ev Idx": 980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967911.432, "dur": 2.526,
    "args": {
      "External id": 1494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967912.359, "dur": 0.989,
    "args": {
      "External id": 1495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967912.843, "dur": 0.363,
    "args": {
      "External id": 1496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967918.191, "dur": 4.943,
    "args": {
      "External id": 1497,"Record function id": 0, "Ev Idx": 984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967920.114, "dur": 2.325,
    "args": {
      "External id": 1498,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967920.979, "dur": 0.921,
    "args": {
      "External id": 1499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967921.360, "dur": 0.379,
    "args": {
      "External id": 1500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967926.542, "dur": 5.257,
    "args": {
      "External id": 1501,"Record function id": 0, "Ev Idx": 988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967928.509, "dur": 2.444,
    "args": {
      "External id": 1502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967929.411, "dur": 1.025,
    "args": {
      "External id": 1503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967929.923, "dur": 0.363,
    "args": {
      "External id": 1504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967935.580, "dur": 5.187,
    "args": {
      "External id": 1505,"Record function id": 0, "Ev Idx": 992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967937.416, "dur": 2.491,
    "args": {
      "External id": 1506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967938.283, "dur": 0.975,
    "args": {
      "External id": 1507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967938.817, "dur": 0.297,
    "args": {
      "External id": 1508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967944.474, "dur": 5.090,
    "args": {
      "External id": 1509,"Record function id": 0, "Ev Idx": 996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967946.302, "dur": 2.397,
    "args": {
      "External id": 1510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967947.264, "dur": 0.855,
    "args": {
      "External id": 1511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967947.684, "dur": 0.299,
    "args": {
      "External id": 1512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967953.307, "dur": 4.811,
    "args": {
      "External id": 1513,"Record function id": 0, "Ev Idx": 1000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967955.243, "dur": 2.165,
    "args": {
      "External id": 1514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967955.995, "dur": 0.876,
    "args": {
      "External id": 1515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967956.380, "dur": 0.344,
    "args": {
      "External id": 1516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967961.478, "dur": 4.471,
    "args": {
      "External id": 1517,"Record function id": 0, "Ev Idx": 1004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967962.961, "dur": 2.285,
    "args": {
      "External id": 1518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967963.710, "dur": 0.985,
    "args": {
      "External id": 1519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967964.073, "dur": 0.493,
    "args": {
      "External id": 1520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967969.340, "dur": 4.942,
    "args": {
      "External id": 1521,"Record function id": 0, "Ev Idx": 1008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967971.005, "dur": 2.557,
    "args": {
      "External id": 1522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967971.916, "dur": 1.129,
    "args": {
      "External id": 1523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967972.441, "dur": 0.461,
    "args": {
      "External id": 1524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967977.696, "dur": 5.196,
    "args": {
      "External id": 1525,"Record function id": 0, "Ev Idx": 1012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967979.564, "dur": 2.499,
    "args": {
      "External id": 1526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967980.528, "dur": 1.008,
    "args": {
      "External id": 1527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967980.919, "dur": 0.492,
    "args": {
      "External id": 1528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967986.277, "dur": 4.871,
    "args": {
      "External id": 1529,"Record function id": 0, "Ev Idx": 1016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967988.093, "dur": 2.346,
    "args": {
      "External id": 1530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967989.031, "dur": 0.867,
    "args": {
      "External id": 1531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967989.445, "dur": 0.327,
    "args": {
      "External id": 1532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967994.536, "dur": 4.775,
    "args": {
      "External id": 1533,"Record function id": 0, "Ev Idx": 1020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507967996.251, "dur": 2.363,
    "args": {
      "External id": 1534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967997.025, "dur": 1.061,
    "args": {
      "External id": 1535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507967997.584, "dur": 0.361,
    "args": {
      "External id": 1536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968253.145, "dur": 20.623,
    "args": {
      "External id": 1537,"Record function id": 0, "Ev Idx": 1024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968268.621, "dur": 4.196,
    "args": {
      "External id": 1538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968270.043, "dur": 2.022,
    "args": {
      "External id": 1539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968271.375, "dur": 0.505,
    "args": {
      "External id": 1540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968278.279, "dur": 5.111,
    "args": {
      "External id": 1541,"Record function id": 0, "Ev Idx": 1028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968280.203, "dur": 2.516,
    "args": {
      "External id": 1542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968281.141, "dur": 1.040,
    "args": {
      "External id": 1543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968281.533, "dur": 0.526,
    "args": {
      "External id": 1544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968286.883, "dur": 4.747,
    "args": {
      "External id": 1545,"Record function id": 0, "Ev Idx": 1032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968288.542, "dur": 2.269,
    "args": {
      "External id": 1546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968289.286, "dur": 0.995,
    "args": {
      "External id": 1547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968289.794, "dur": 0.362,
    "args": {
      "External id": 1548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968295.330, "dur": 5.254,
    "args": {
      "External id": 1549,"Record function id": 0, "Ev Idx": 1036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968297.328, "dur": 2.412,
    "args": {
      "External id": 1550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968298.256, "dur": 0.810,
    "args": {
      "External id": 1551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968298.608, "dur": 0.328,
    "args": {
      "External id": 1552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968304.399, "dur": 4.854,
    "args": {
      "External id": 1553,"Record function id": 0, "Ev Idx": 1040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968306.057, "dur": 2.490,
    "args": {
      "External id": 1554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968307.089, "dur": 0.797,
    "args": {
      "External id": 1555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968307.457, "dur": 0.294,
    "args": {
      "External id": 1556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968312.415, "dur": 5.091,
    "args": {
      "External id": 1557,"Record function id": 0, "Ev Idx": 1044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968314.283, "dur": 2.535,
    "args": {
      "External id": 1558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968315.038, "dur": 1.275,
    "args": {
      "External id": 1559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968315.428, "dur": 0.754,
    "args": {
      "External id": 1560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968321.186, "dur": 4.726,
    "args": {
      "External id": 1561,"Record function id": 0, "Ev Idx": 1048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968323.155, "dur": 2.063,
    "args": {
      "External id": 1562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968323.906, "dur": 0.819,
    "args": {
      "External id": 1563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968324.255, "dur": 0.338,
    "args": {
      "External id": 1564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968329.513, "dur": 5.544,
    "args": {
      "External id": 1565,"Record function id": 0, "Ev Idx": 1052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968331.239, "dur": 3.086,
    "args": {
      "External id": 1566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968332.966, "dur": 0.806,
    "args": {
      "External id": 1567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968333.343, "dur": 0.300,
    "args": {
      "External id": 1568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968338.675, "dur": 54.545,
    "args": {
      "External id": 1569,"Record function id": 0, "Ev Idx": 1056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968340.406, "dur": 48.474,
    "args": {
      "External id": 1570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968341.069, "dur": 47.019,
    "args": {
      "External id": 1571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968341.401, "dur": 45.343,
    "args": {
      "External id": 1572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968401.483, "dur": 8.898,
    "args": {
      "External id": 1573,"Record function id": 0, "Ev Idx": 1060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968407.197, "dur": 2.364,
    "args": {
      "External id": 1574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968408.070, "dur": 0.976,
    "args": {
      "External id": 1575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968408.526, "dur": 0.391,
    "args": {
      "External id": 1576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968413.780, "dur": 4.860,
    "args": {
      "External id": 1577,"Record function id": 0, "Ev Idx": 1064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968415.697, "dur": 2.250,
    "args": {
      "External id": 1578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968416.423, "dur": 1.053,
    "args": {
      "External id": 1579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968416.818, "dur": 0.534,
    "args": {
      "External id": 1580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968422.008, "dur": 6.029,
    "args": {
      "External id": 1581,"Record function id": 0, "Ev Idx": 1068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968424.319, "dur": 2.926,
    "args": {
      "External id": 1582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968425.436, "dur": 0.847,
    "args": {
      "External id": 1583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968425.790, "dur": 0.341,
    "args": {
      "External id": 1584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968431.447, "dur": 5.091,
    "args": {
      "External id": 1585,"Record function id": 0, "Ev Idx": 1072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968433.530, "dur": 2.307,
    "args": {
      "External id": 1586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968434.357, "dur": 0.914,
    "args": {
      "External id": 1587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968434.686, "dur": 0.457,
    "args": {
      "External id": 1588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968439.839, "dur": 4.681,
    "args": {
      "External id": 1589,"Record function id": 0, "Ev Idx": 1076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968441.488, "dur": 2.170,
    "args": {
      "External id": 1590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968442.189, "dur": 1.027,
    "args": {
      "External id": 1591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968442.746, "dur": 0.334,
    "args": {
      "External id": 1592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968447.881, "dur": 5.672,
    "args": {
      "External id": 1593,"Record function id": 0, "Ev Idx": 1080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968450.505, "dur": 2.333,
    "args": {
      "External id": 1594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968451.294, "dur": 0.929,
    "args": {
      "External id": 1595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968451.626, "dur": 0.470,
    "args": {
      "External id": 1596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968456.926, "dur": 4.755,
    "args": {
      "External id": 1597,"Record function id": 0, "Ev Idx": 1084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968458.882, "dur": 2.103,
    "args": {
      "External id": 1598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968459.562, "dur": 0.956,
    "args": {
      "External id": 1599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968459.944, "dur": 0.402,
    "args": {
      "External id": 1600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968465.204, "dur": 4.838,
    "args": {
      "External id": 1601,"Record function id": 0, "Ev Idx": 1088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968467.058, "dur": 2.296,
    "args": {
      "External id": 1602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968467.888, "dur": 0.971,
    "args": {
      "External id": 1603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968468.226, "dur": 0.503,
    "args": {
      "External id": 1604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968473.474, "dur": 4.590,
    "args": {
      "External id": 1605,"Record function id": 0, "Ev Idx": 1092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968475.298, "dur": 2.060,
    "args": {
      "External id": 1606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968476.091, "dur": 0.807,
    "args": {
      "External id": 1607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968476.468, "dur": 0.301,
    "args": {
      "External id": 1608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968481.396, "dur": 4.530,
    "args": {
      "External id": 1609,"Record function id": 0, "Ev Idx": 1096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968482.944, "dur": 2.122,
    "args": {
      "External id": 1610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968483.678, "dur": 0.892,
    "args": {
      "External id": 1611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968484.115, "dur": 0.293,
    "args": {
      "External id": 1612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968489.275, "dur": 4.689,
    "args": {
      "External id": 1613,"Record function id": 0, "Ev Idx": 1100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968491.025, "dur": 2.116,
    "args": {
      "External id": 1614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968491.842, "dur": 0.845,
    "args": {
      "External id": 1615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968492.247, "dur": 0.306,
    "args": {
      "External id": 1616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968497.307, "dur": 4.707,
    "args": {
      "External id": 1617,"Record function id": 0, "Ev Idx": 1104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968499.018, "dur": 2.330,
    "args": {
      "External id": 1618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968499.864, "dur": 0.986,
    "args": {
      "External id": 1619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968500.360, "dur": 0.359,
    "args": {
      "External id": 1620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968506.741, "dur": 7.007,
    "args": {
      "External id": 1621,"Record function id": 0, "Ev Idx": 1108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968509.839, "dur": 3.211,
    "args": {
      "External id": 1622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968510.689, "dur": 1.874,
    "args": {
      "External id": 1623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968511.328, "dur": 1.109,
    "args": {
      "External id": 1624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968517.375, "dur": 4.742,
    "args": {
      "External id": 1625,"Record function id": 0, "Ev Idx": 1112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968519.159, "dur": 2.261,
    "args": {
      "External id": 1626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968519.924, "dur": 0.980,
    "args": {
      "External id": 1627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968520.332, "dur": 0.437,
    "args": {
      "External id": 1628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968525.565, "dur": 4.374,
    "args": {
      "External id": 1629,"Record function id": 0, "Ev Idx": 1116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968527.161, "dur": 1.952,
    "args": {
      "External id": 1630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968527.846, "dur": 0.798,
    "args": {
      "External id": 1631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968528.181, "dur": 0.329,
    "args": {
      "External id": 1632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968533.449, "dur": 5.346,
    "args": {
      "External id": 1633,"Record function id": 0, "Ev Idx": 1120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968535.721, "dur": 2.348,
    "args": {
      "External id": 1634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968536.590, "dur": 0.872,
    "args": {
      "External id": 1635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968536.920, "dur": 0.405,
    "args": {
      "External id": 1636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968542.189, "dur": 5.220,
    "args": {
      "External id": 1637,"Record function id": 0, "Ev Idx": 1124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968543.982, "dur": 2.628,
    "args": {
      "External id": 1638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968544.834, "dur": 1.188,
    "args": {
      "External id": 1639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968545.165, "dur": 0.719,
    "args": {
      "External id": 1640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968550.818, "dur": 4.221,
    "args": {
      "External id": 1641,"Record function id": 0, "Ev Idx": 1128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968552.312, "dur": 2.016,
    "args": {
      "External id": 1642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968552.935, "dur": 1.091,
    "args": {
      "External id": 1643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968553.314, "dur": 0.569,
    "args": {
      "External id": 1644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968558.492, "dur": 5.521,
    "args": {
      "External id": 1645,"Record function id": 0, "Ev Idx": 1132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968560.148, "dur": 3.018,
    "args": {
      "External id": 1646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968560.903, "dur": 1.760,
    "args": {
      "External id": 1647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968562.053, "dur": 0.484,
    "args": {
      "External id": 1648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968567.529, "dur": 4.664,
    "args": {
      "External id": 1649,"Record function id": 0, "Ev Idx": 1136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968569.188, "dur": 2.311,
    "args": {
      "External id": 1650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968569.940, "dur": 1.058,
    "args": {
      "External id": 1651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968570.339, "dur": 0.512,
    "args": {
      "External id": 1652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968575.873, "dur": 55.325,
    "args": {
      "External id": 1653,"Record function id": 0, "Ev Idx": 1140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968577.517, "dur": 52.317,
    "args": {
      "External id": 1654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968578.282, "dur": 50.777,
    "args": {
      "External id": 1655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968578.619, "dur": 48.399,
    "args": {
      "External id": 1656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968642.276, "dur": 11.043,
    "args": {
      "External id": 1657,"Record function id": 0, "Ev Idx": 1144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968650.010, "dur": 2.447,
    "args": {
      "External id": 1658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968651.009, "dur": 0.971,
    "args": {
      "External id": 1659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968651.462, "dur": 0.361,
    "args": {
      "External id": 1660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968656.702, "dur": 4.856,
    "args": {
      "External id": 1661,"Record function id": 0, "Ev Idx": 1148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968658.720, "dur": 2.143,
    "args": {
      "External id": 1662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968659.497, "dur": 0.900,
    "args": {
      "External id": 1663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968659.867, "dur": 0.382,
    "args": {
      "External id": 1664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968664.900, "dur": 5.221,
    "args": {
      "External id": 1665,"Record function id": 0, "Ev Idx": 1152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968666.776, "dur": 2.512,
    "args": {
      "External id": 1666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968667.683, "dur": 1.110,
    "args": {
      "External id": 1667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968668.039, "dur": 0.619,
    "args": {
      "External id": 1668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968673.506, "dur": 66.055,
    "args": {
      "External id": 1669,"Record function id": 0, "Ev Idx": 1156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968675.088, "dur": 63.079,
    "args": {
      "External id": 1670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968735.607, "dur": 1.800,
    "args": {
      "External id": 1671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968736.608, "dur": 0.614,
    "args": {
      "External id": 1672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968746.383, "dur": 6.523,
    "args": {
      "External id": 1673,"Record function id": 0, "Ev Idx": 1160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968749.451, "dur": 2.617,
    "args": {
      "External id": 1674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968750.580, "dur": 1.013,
    "args": {
      "External id": 1675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968751.033, "dur": 0.435,
    "args": {
      "External id": 1676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968756.737, "dur": 5.246,
    "args": {
      "External id": 1677,"Record function id": 0, "Ev Idx": 1164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968758.813, "dur": 2.333,
    "args": {
      "External id": 1678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968759.580, "dur": 1.050,
    "args": {
      "External id": 1679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968760.051, "dur": 0.445,
    "args": {
      "External id": 1680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968765.614, "dur": 5.285,
    "args": {
      "External id": 1681,"Record function id": 0, "Ev Idx": 1168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968767.794, "dur": 2.260,
    "args": {
      "External id": 1682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968768.680, "dur": 0.914,
    "args": {
      "External id": 1683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968769.149, "dur": 0.317,
    "args": {
      "External id": 1684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968774.334, "dur": 4.672,
    "args": {
      "External id": 1685,"Record function id": 0, "Ev Idx": 1172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968776.133, "dur": 2.029,
    "args": {
      "External id": 1686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968776.851, "dur": 0.842,
    "args": {
      "External id": 1687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968777.196, "dur": 0.370,
    "args": {
      "External id": 1688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968782.285, "dur": 4.178,
    "args": {
      "External id": 1689,"Record function id": 0, "Ev Idx": 1176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968783.715, "dur": 2.085,
    "args": {
      "External id": 1690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968784.432, "dur": 0.863,
    "args": {
      "External id": 1691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968784.805, "dur": 0.349,
    "args": {
      "External id": 1692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968789.630, "dur": 4.830,
    "args": {
      "External id": 1693,"Record function id": 0, "Ev Idx": 1180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968791.355, "dur": 2.293,
    "args": {
      "External id": 1694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968792.122, "dur": 1.045,
    "args": {
      "External id": 1695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968792.517, "dur": 0.470,
    "args": {
      "External id": 1696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968797.917, "dur": 4.634,
    "args": {
      "External id": 1697,"Record function id": 0, "Ev Idx": 1184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968799.591, "dur": 2.123,
    "args": {
      "External id": 1698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 1185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968800.319, "dur": 0.900,
    "args": {
      "External id": 1699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 1186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968800.681, "dur": 0.408,
    "args": {
      "External id": 1700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 1187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968809.280, "dur": 5.349,
    "args": {
      "External id": 1701,"Record function id": 0, "Ev Idx": 1188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968811.528, "dur": 2.304,
    "args": {
      "External id": 1702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968812.379, "dur": 0.954,
    "args": {
      "External id": 1703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968812.821, "dur": 0.335,
    "args": {
      "External id": 1704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968818.000, "dur": 4.909,
    "args": {
      "External id": 1705,"Record function id": 0, "Ev Idx": 1192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968819.631, "dur": 2.451,
    "args": {
      "External id": 1706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968820.475, "dur": 0.983,
    "args": {
      "External id": 1707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968820.837, "dur": 0.476,
    "args": {
      "External id": 1708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968826.291, "dur": 5.245,
    "args": {
      "External id": 1709,"Record function id": 0, "Ev Idx": 1196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968828.031, "dur": 2.827,
    "args": {
      "External id": 1710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968828.868, "dur": 1.376,
    "args": {
      "External id": 1711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968829.238, "dur": 0.871,
    "args": {
      "External id": 1712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968835.023, "dur": 5.201,
    "args": {
      "External id": 1713,"Record function id": 0, "Ev Idx": 1200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968837.188, "dur": 2.204,
    "args": {
      "External id": 1714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968837.969, "dur": 0.953,
    "args": {
      "External id": 1715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968838.336, "dur": 0.455,
    "args": {
      "External id": 1716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968843.540, "dur": 4.885,
    "args": {
      "External id": 1717,"Record function id": 0, "Ev Idx": 1204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968845.468, "dur": 2.253,
    "args": {
      "External id": 1718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968846.310, "dur": 0.949,
    "args": {
      "External id": 1719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968846.667, "dur": 0.462,
    "args": {
      "External id": 1720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968851.844, "dur": 4.632,
    "args": {
      "External id": 1721,"Record function id": 0, "Ev Idx": 1208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968853.486, "dur": 2.295,
    "args": {
      "External id": 1722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968854.331, "dur": 0.945,
    "args": {
      "External id": 1723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968854.716, "dur": 0.429,
    "args": {
      "External id": 1724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968860.161, "dur": 4.487,
    "args": {
      "External id": 1725,"Record function id": 0, "Ev Idx": 1212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968861.936, "dur": 2.024,
    "args": {
      "External id": 1726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968862.669, "dur": 0.819,
    "args": {
      "External id": 1727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968863.028, "dur": 0.323,
    "args": {
      "External id": 1728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968867.929, "dur": 4.868,
    "args": {
      "External id": 1729,"Record function id": 0, "Ev Idx": 1216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968869.707, "dur": 2.378,
    "args": {
      "External id": 1730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968870.430, "dur": 1.167,
    "args": {
      "External id": 1731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968870.803, "dur": 0.640,
    "args": {
      "External id": 1732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968876.066, "dur": 5.494,
    "args": {
      "External id": 1733,"Record function id": 0, "Ev Idx": 1220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968878.173, "dur": 2.569,
    "args": {
      "External id": 1734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968879.131, "dur": 1.126,
    "args": {
      "External id": 1735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968879.502, "dur": 0.607,
    "args": {
      "External id": 1736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968885.121, "dur": 56.510,
    "args": {
      "External id": 1737,"Record function id": 0, "Ev Idx": 1224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968886.851, "dur": 50.415,
    "args": {
      "External id": 1738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968887.793, "dur": 48.676,
    "args": {
      "External id": 1739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968888.148, "dur": 46.946,
    "args": {
      "External id": 1740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968948.925, "dur": 9.670,
    "args": {
      "External id": 1741,"Record function id": 0, "Ev Idx": 1228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968955.252, "dur": 2.491,
    "args": {
      "External id": 1742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968956.162, "dur": 1.095,
    "args": {
      "External id": 1743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968956.674, "dur": 0.458,
    "args": {
      "External id": 1744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968962.143, "dur": 5.288,
    "args": {
      "External id": 1745,"Record function id": 0, "Ev Idx": 1232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968963.826, "dur": 2.777,
    "args": {
      "External id": 1746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968964.841, "dur": 1.279,
    "args": {
      "External id": 1747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968965.501, "dur": 0.474,
    "args": {
      "External id": 1748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968970.809, "dur": 5.789,
    "args": {
      "External id": 1749,"Record function id": 0, "Ev Idx": 1236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968972.981, "dur": 2.749,
    "args": {
      "External id": 1750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968974.025, "dur": 0.981,
    "args": {
      "External id": 1751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968974.549, "dur": 0.296,
    "args": {
      "External id": 1752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968979.996, "dur": 6.298,
    "args": {
      "External id": 1753,"Record function id": 0, "Ev Idx": 1240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968982.182, "dur": 3.352,
    "args": {
      "External id": 1754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968983.919, "dur": 1.011,
    "args": {
      "External id": 1755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968984.484, "dur": 0.320,
    "args": {
      "External id": 1756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968989.694, "dur": 4.473,
    "args": {
      "External id": 1757,"Record function id": 0, "Ev Idx": 1244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968991.209, "dur": 2.210,
    "args": {
      "External id": 1758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968992.063, "dur": 0.916,
    "args": {
      "External id": 1759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968992.438, "dur": 0.411,
    "args": {
      "External id": 1760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968997.522, "dur": 4.505,
    "args": {
      "External id": 1761,"Record function id": 0, "Ev Idx": 1248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507968999.041, "dur": 2.137,
    "args": {
      "External id": 1762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507968999.831, "dur": 0.882,
    "args": {
      "External id": 1763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969000.224, "dur": 0.366,
    "args": {
      "External id": 1764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969005.318, "dur": 4.797,
    "args": {
      "External id": 1765,"Record function id": 0, "Ev Idx": 1252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969007.096, "dur": 2.198,
    "args": {
      "External id": 1766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969007.869, "dur": 0.939,
    "args": {
      "External id": 1767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969008.245, "dur": 0.429,
    "args": {
      "External id": 1768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969013.682, "dur": 4.719,
    "args": {
      "External id": 1769,"Record function id": 0, "Ev Idx": 1256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969015.311, "dur": 2.278,
    "args": {
      "External id": 1770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969016.057, "dur": 1.041,
    "args": {
      "External id": 1771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969016.421, "dur": 0.545,
    "args": {
      "External id": 1772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969022.117, "dur": 4.884,
    "args": {
      "External id": 1773,"Record function id": 0, "Ev Idx": 1260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969024.037, "dur": 2.180,
    "args": {
      "External id": 1774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969024.783, "dur": 0.963,
    "args": {
      "External id": 1775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969025.176, "dur": 0.439,
    "args": {
      "External id": 1776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969030.363, "dur": 4.666,
    "args": {
      "External id": 1777,"Record function id": 0, "Ev Idx": 1264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969032.111, "dur": 2.218,
    "args": {
      "External id": 1778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969032.903, "dur": 0.945,
    "args": {
      "External id": 1779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969033.274, "dur": 0.443,
    "args": {
      "External id": 1780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969038.350, "dur": 5.586,
    "args": {
      "External id": 1781,"Record function id": 0, "Ev Idx": 1268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969040.805, "dur": 2.283,
    "args": {
      "External id": 1782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969041.612, "dur": 0.995,
    "args": {
      "External id": 1783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969042.014, "dur": 0.463,
    "args": {
      "External id": 1784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969047.435, "dur": 4.709,
    "args": {
      "External id": 1785,"Record function id": 0, "Ev Idx": 1272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969049.121, "dur": 2.215,
    "args": {
      "External id": 1786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969049.911, "dur": 0.950,
    "args": {
      "External id": 1787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969050.284, "dur": 0.452,
    "args": {
      "External id": 1788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969055.503, "dur": 5.373,
    "args": {
      "External id": 1789,"Record function id": 0, "Ev Idx": 1276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969057.460, "dur": 2.684,
    "args": {
      "External id": 1790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969058.286, "dur": 1.358,
    "args": {
      "External id": 1791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969058.643, "dur": 0.877,
    "args": {
      "External id": 1792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969064.187, "dur": 4.830,
    "args": {
      "External id": 1793,"Record function id": 0, "Ev Idx": 1280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969065.944, "dur": 2.320,
    "args": {
      "External id": 1794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969066.893, "dur": 0.839,
    "args": {
      "External id": 1795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969067.284, "dur": 0.313,
    "args": {
      "External id": 1796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969072.386, "dur": 4.970,
    "args": {
      "External id": 1797,"Record function id": 0, "Ev Idx": 1284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969074.064, "dur": 2.377,
    "args": {
      "External id": 1798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969074.945, "dur": 0.869,
    "args": {
      "External id": 1799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969075.310, "dur": 0.368,
    "args": {
      "External id": 1800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969084.307, "dur": 4.793,
    "args": {
      "External id": 1801,"Record function id": 0, "Ev Idx": 1288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969085.998, "dur": 2.259,
    "args": {
      "External id": 1802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969086.839, "dur": 0.942,
    "args": {
      "External id": 1803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969087.314, "dur": 0.329,
    "args": {
      "External id": 1804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969092.563, "dur": 5.614,
    "args": {
      "External id": 1805,"Record function id": 0, "Ev Idx": 1292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969094.722, "dur": 2.640,
    "args": {
      "External id": 1806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969095.538, "dur": 1.296,
    "args": {
      "External id": 1807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969095.993, "dur": 0.694,
    "args": {
      "External id": 1808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969102.523, "dur": 5.088,
    "args": {
      "External id": 1809,"Record function id": 0, "Ev Idx": 1296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969104.543, "dur": 2.358,
    "args": {
      "External id": 1810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969105.323, "dur": 1.125,
    "args": {
      "External id": 1811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969105.787, "dur": 0.520,
    "args": {
      "External id": 1812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969110.905, "dur": 5.325,
    "args": {
      "External id": 1813,"Record function id": 0, "Ev Idx": 1300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969112.987, "dur": 2.376,
    "args": {
      "External id": 1814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969113.810, "dur": 1.101,
    "args": {
      "External id": 1815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969114.283, "dur": 0.500,
    "args": {
      "External id": 1816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969120.362, "dur": 4.933,
    "args": {
      "External id": 1817,"Record function id": 0, "Ev Idx": 1304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969122.217, "dur": 2.379,
    "args": {
      "External id": 1818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969122.902, "dur": 1.229,
    "args": {
      "External id": 1819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969123.304, "dur": 0.701,
    "args": {
      "External id": 1820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969128.610, "dur": 57.215,
    "args": {
      "External id": 1821,"Record function id": 0, "Ev Idx": 1308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969130.007, "dur": 54.449,
    "args": {
      "External id": 1822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969130.734, "dur": 52.958,
    "args": {
      "External id": 1823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969131.143, "dur": 51.200,
    "args": {
      "External id": 1824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969199.569, "dur": 8.683,
    "args": {
      "External id": 1825,"Record function id": 0, "Ev Idx": 1312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969205.097, "dur": 2.449,
    "args": {
      "External id": 1826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969205.974, "dur": 1.075,
    "args": {
      "External id": 1827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969206.407, "dur": 0.493,
    "args": {
      "External id": 1828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969211.520, "dur": 4.298,
    "args": {
      "External id": 1829,"Record function id": 0, "Ev Idx": 1316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969213.062, "dur": 2.058,
    "args": {
      "External id": 1830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969213.822, "dur": 0.831,
    "args": {
      "External id": 1831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969214.198, "dur": 0.312,
    "args": {
      "External id": 1832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969219.161, "dur": 64.697,
    "args": {
      "External id": 1833,"Record function id": 0, "Ev Idx": 1320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969221.295, "dur": 57.875,
    "args": {
      "External id": 1834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969222.154, "dur": 56.166,
    "args": {
      "External id": 1835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969277.140, "dur": 0.945,
    "args": {
      "External id": 1836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969288.305, "dur": 8.337,
    "args": {
      "External id": 1837,"Record function id": 0, "Ev Idx": 1324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969293.340, "dur": 2.420,
    "args": {
      "External id": 1838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969294.220, "dur": 1.052,
    "args": {
      "External id": 1839,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969294.622, "dur": 0.521,
    "args": {
      "External id": 1840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969300.180, "dur": 5.325,
    "args": {
      "External id": 1841,"Record function id": 0, "Ev Idx": 1328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969302.187, "dur": 2.468,
    "args": {
      "External id": 1842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969303.120, "dur": 0.878,
    "args": {
      "External id": 1843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969303.516, "dur": 0.347,
    "args": {
      "External id": 1844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969308.936, "dur": 4.796,
    "args": {
      "External id": 1845,"Record function id": 0, "Ev Idx": 1332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969310.570, "dur": 2.354,
    "args": {
      "External id": 1846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969311.454, "dur": 1.005,
    "args": {
      "External id": 1847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969311.914, "dur": 0.413,
    "args": {
      "External id": 1848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969317.116, "dur": 4.562,
    "args": {
      "External id": 1849,"Record function id": 0, "Ev Idx": 1336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969318.673, "dur": 2.160,
    "args": {
      "External id": 1850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969319.489, "dur": 0.838,
    "args": {
      "External id": 1851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969319.854, "dur": 0.347,
    "args": {
      "External id": 1852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969325.027, "dur": 4.818,
    "args": {
      "External id": 1853,"Record function id": 0, "Ev Idx": 1340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969326.805, "dur": 2.228,
    "args": {
      "External id": 1854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969327.647, "dur": 0.930,
    "args": {
      "External id": 1855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969328.008, "dur": 0.438,
    "args": {
      "External id": 1856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969333.398, "dur": 5.169,
    "args": {
      "External id": 1857,"Record function id": 0, "Ev Idx": 1344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969335.381, "dur": 2.346,
    "args": {
      "External id": 1858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969336.257, "dur": 0.980,
    "args": {
      "External id": 1859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969336.651, "dur": 0.447,
    "args": {
      "External id": 1860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969341.864, "dur": 4.587,
    "args": {
      "External id": 1861,"Record function id": 0, "Ev Idx": 1348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969343.317, "dur": 2.334,
    "args": {
      "External id": 1862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969344.234, "dur": 0.930,
    "args": {
      "External id": 1863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969344.575, "dur": 0.456,
    "args": {
      "External id": 1864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969349.711, "dur": 4.948,
    "args": {
      "External id": 1865,"Record function id": 0, "Ev Idx": 1352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969351.639, "dur": 2.338,
    "args": {
      "External id": 1866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969352.453, "dur": 1.050,
    "args": {
      "External id": 1867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969352.884, "dur": 0.472,
    "args": {
      "External id": 1868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969358.014, "dur": 4.642,
    "args": {
      "External id": 1869,"Record function id": 0, "Ev Idx": 1356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969359.686, "dur": 2.180,
    "args": {
      "External id": 1870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969360.415, "dur": 0.977,
    "args": {
      "External id": 1871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969360.772, "dur": 0.478,
    "args": {
      "External id": 1872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969366.151, "dur": 4.835,
    "args": {
      "External id": 1873,"Record function id": 0, "Ev Idx": 1360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969368.152, "dur": 2.165,
    "args": {
      "External id": 1874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969369.018, "dur": 0.815,
    "args": {
      "External id": 1875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969369.378, "dur": 0.326,
    "args": {
      "External id": 1876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969374.468, "dur": 82.858,
    "args": {
      "External id": 1877,"Record function id": 0, "Ev Idx": 1364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969376.998, "dur": 78.919,
    "args": {
      "External id": 1878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969453.418, "dur": 1.672,
    "args": {
      "External id": 1879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969454.304, "dur": 0.599,
    "args": {
      "External id": 1880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969461.563, "dur": 5.772,
    "args": {
      "External id": 1881,"Record function id": 0, "Ev Idx": 1368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969463.881, "dur": 2.749,
    "args": {
      "External id": 1882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969464.905, "dur": 1.090,
    "args": {
      "External id": 1883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969465.268, "dur": 0.580,
    "args": {
      "External id": 1884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969470.895, "dur": 4.307,
    "args": {
      "External id": 1885,"Record function id": 0, "Ev Idx": 1372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969472.402, "dur": 2.120,
    "args": {
      "External id": 1886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969473.186, "dur": 0.857,
    "args": {
      "External id": 1887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969473.558, "dur": 0.353,
    "args": {
      "External id": 1888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969478.855, "dur": 4.834,
    "args": {
      "External id": 1889,"Record function id": 0, "Ev Idx": 1376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969480.803, "dur": 2.072,
    "args": {
      "External id": 1890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969481.571, "dur": 0.836,
    "args": {
      "External id": 1891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969481.926, "dur": 0.347,
    "args": {
      "External id": 1892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969488.850, "dur": 4.795,
    "args": {
      "External id": 1893,"Record function id": 0, "Ev Idx": 1380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969490.894, "dur": 2.047,
    "args": {
      "External id": 1894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969491.617, "dur": 0.838,
    "args": {
      "External id": 1895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969491.971, "dur": 0.330,
    "args": {
      "External id": 1896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969497.227, "dur": 4.620,
    "args": {
      "External id": 1897,"Record function id": 0, "Ev Idx": 1384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969498.875, "dur": 2.149,
    "args": {
      "External id": 1898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969499.687, "dur": 0.842,
    "args": {
      "External id": 1899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969500.087, "dur": 0.314,
    "args": {
      "External id": 1900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969505.218, "dur": 4.667,
    "args": {
      "External id": 1901,"Record function id": 0, "Ev Idx": 1388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969507.094, "dur": 2.103,
    "args": {
      "External id": 1902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969507.727, "dur": 0.997,
    "args": {
      "External id": 1903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969508.066, "dur": 0.525,
    "args": {
      "External id": 1904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969513.237, "dur": 4.239,
    "args": {
      "External id": 1905,"Record function id": 0, "Ev Idx": 1392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969514.655, "dur": 2.139,
    "args": {
      "External id": 1906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969515.406, "dur": 0.930,
    "args": {
      "External id": 1907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969515.771, "dur": 0.435,
    "args": {
      "External id": 1908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969520.826, "dur": 4.541,
    "args": {
      "External id": 1909,"Record function id": 0, "Ev Idx": 1396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969522.222, "dur": 2.286,
    "args": {
      "External id": 1910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969523.051, "dur": 0.989,
    "args": {
      "External id": 1911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969523.405, "dur": 0.508,
    "args": {
      "External id": 1912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969528.800, "dur": 4.204,
    "args": {
      "External id": 1913,"Record function id": 0, "Ev Idx": 1400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969530.230, "dur": 2.086,
    "args": {
      "External id": 1914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969530.946, "dur": 0.913,
    "args": {
      "External id": 1915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969531.286, "dur": 0.449,
    "args": {
      "External id": 1916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969536.546, "dur": 4.513,
    "args": {
      "External id": 1917,"Record function id": 0, "Ev Idx": 1404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969538.269, "dur": 2.126,
    "args": {
      "External id": 1918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969538.994, "dur": 0.915,
    "args": {
      "External id": 1919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969539.349, "dur": 0.427,
    "args": {
      "External id": 1920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969544.458, "dur": 5.175,
    "args": {
      "External id": 1921,"Record function id": 0, "Ev Idx": 1408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969546.583, "dur": 2.359,
    "args": {
      "External id": 1922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969547.483, "dur": 0.842,
    "args": {
      "External id": 1923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969547.845, "dur": 0.333,
    "args": {
      "External id": 1924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969552.847, "dur": 5.328,
    "args": {
      "External id": 1925,"Record function id": 0, "Ev Idx": 1412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969554.946, "dur": 2.535,
    "args": {
      "External id": 1926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969555.862, "dur": 0.978,
    "args": {
      "External id": 1927,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969556.222, "dur": 0.480,
    "args": {
      "External id": 1928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969561.335, "dur": 4.300,
    "args": {
      "External id": 1929,"Record function id": 0, "Ev Idx": 1416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969562.809, "dur": 2.137,
    "args": {
      "External id": 1930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969563.497, "dur": 0.962,
    "args": {
      "External id": 1931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969563.913, "dur": 0.418,
    "args": {
      "External id": 1932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969568.821, "dur": 4.419,
    "args": {
      "External id": 1933,"Record function id": 0, "Ev Idx": 1420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969570.363, "dur": 2.125,
    "args": {
      "External id": 1934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969571.076, "dur": 0.950,
    "args": {
      "External id": 1935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969571.431, "dur": 0.460,
    "args": {
      "External id": 1936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969576.395, "dur": 4.714,
    "args": {
      "External id": 1937,"Record function id": 0, "Ev Idx": 1424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969577.950, "dur": 2.322,
    "args": {
      "External id": 1938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969578.825, "dur": 0.962,
    "args": {
      "External id": 1939,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969579.195, "dur": 0.456,
    "args": {
      "External id": 1940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969584.247, "dur": 5.661,
    "args": {
      "External id": 1941,"Record function id": 0, "Ev Idx": 1428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969585.934, "dur": 3.260,
    "args": {
      "External id": 1942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969587.038, "dur": 1.671,
    "args": {
      "External id": 1943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969587.399, "dur": 1.183,
    "args": {
      "External id": 1944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969594.134, "dur": 4.439,
    "args": {
      "External id": 1945,"Record function id": 0, "Ev Idx": 1432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969595.724, "dur": 2.159,
    "args": {
      "External id": 1946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969596.459, "dur": 0.957,
    "args": {
      "External id": 1947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969596.826, "dur": 0.460,
    "args": {
      "External id": 1948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969601.708, "dur": 4.709,
    "args": {
      "External id": 1949,"Record function id": 0, "Ev Idx": 1436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969603.458, "dur": 2.255,
    "args": {
      "External id": 1950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969604.211, "dur": 1.032,
    "args": {
      "External id": 1951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969604.592, "dur": 0.496,
    "args": {
      "External id": 1952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969609.864, "dur": 4.507,
    "args": {
      "External id": 1953,"Record function id": 0, "Ev Idx": 1440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969611.553, "dur": 2.116,
    "args": {
      "External id": 1954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969612.224, "dur": 0.985,
    "args": {
      "External id": 1955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969612.588, "dur": 0.487,
    "args": {
      "External id": 1956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969617.907, "dur": 5.452,
    "args": {
      "External id": 1957,"Record function id": 0, "Ev Idx": 1444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969620.154, "dur": 2.377,
    "args": {
      "External id": 1958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969620.914, "dur": 1.085,
    "args": {
      "External id": 1959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969621.273, "dur": 0.578,
    "args": {
      "External id": 1960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969627.046, "dur": 5.375,
    "args": {
      "External id": 1961,"Record function id": 0, "Ev Idx": 1448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969629.080, "dur": 2.658,
    "args": {
      "External id": 1962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969629.970, "dur": 0.986,
    "args": {
      "External id": 1963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969630.334, "dur": 0.494,
    "args": {
      "External id": 1964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969635.897, "dur": 5.581,
    "args": {
      "External id": 1965,"Record function id": 0, "Ev Idx": 1452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969637.797, "dur": 2.826,
    "args": {
      "External id": 1966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969638.722, "dur": 1.294,
    "args": {
      "External id": 1967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969639.075, "dur": 0.769,
    "args": {
      "External id": 1968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969645.096, "dur": 4.775,
    "args": {
      "External id": 1969,"Record function id": 0, "Ev Idx": 1456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969647.071, "dur": 1.989,
    "args": {
      "External id": 1970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969647.739, "dur": 0.816,
    "args": {
      "External id": 1971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969648.078, "dur": 0.339,
    "args": {
      "External id": 1972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969653.423, "dur": 53.138,
    "args": {
      "External id": 1973,"Record function id": 0, "Ev Idx": 1460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969655.331, "dur": 49.895,
    "args": {
      "External id": 1974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969656.205, "dur": 48.198,
    "args": {
      "External id": 1975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969656.566, "dur": 46.393,
    "args": {
      "External id": 1976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969721.149, "dur": 5.012,
    "args": {
      "External id": 1977,"Record function id": 0, "Ev Idx": 1464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969722.976, "dur": 2.465,
    "args": {
      "External id": 1978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969723.836, "dur": 1.071,
    "args": {
      "External id": 1979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969724.279, "dur": 0.496,
    "args": {
      "External id": 1980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969729.653, "dur": 4.757,
    "args": {
      "External id": 1981,"Record function id": 0, "Ev Idx": 1468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969731.569, "dur": 2.164,
    "args": {
      "External id": 1982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969732.341, "dur": 0.860,
    "args": {
      "External id": 1983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969732.754, "dur": 0.312,
    "args": {
      "External id": 1984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969737.946, "dur": 4.954,
    "args": {
      "External id": 1985,"Record function id": 0, "Ev Idx": 1472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969739.848, "dur": 2.230,
    "args": {
      "External id": 1986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969740.603, "dur": 0.947,
    "args": {
      "External id": 1987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969740.966, "dur": 0.454,
    "args": {
      "External id": 1988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969746.192, "dur": 4.599,
    "args": {
      "External id": 1989,"Record function id": 0, "Ev Idx": 1476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969747.825, "dur": 2.156,
    "args": {
      "External id": 1990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969748.513, "dur": 0.948,
    "args": {
      "External id": 1991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969748.896, "dur": 0.438,
    "args": {
      "External id": 1992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969754.250, "dur": 4.580,
    "args": {
      "External id": 1993,"Record function id": 0, "Ev Idx": 1480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969755.714, "dur": 2.297,
    "args": {
      "External id": 1994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969756.584, "dur": 0.912,
    "args": {
      "External id": 1995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969756.964, "dur": 0.405,
    "args": {
      "External id": 1996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969762.461, "dur": 4.587,
    "args": {
      "External id": 1997,"Record function id": 0, "Ev Idx": 1484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969763.970, "dur": 2.235,
    "args": {
      "External id": 1998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969764.730, "dur": 0.935,
    "args": {
      "External id": 1999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969765.098, "dur": 0.433,
    "args": {
      "External id": 2000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969770.026, "dur": 4.491,
    "args": {
      "External id": 2001,"Record function id": 0, "Ev Idx": 1488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969771.574, "dur": 2.276,
    "args": {
      "External id": 2002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969772.372, "dur": 0.967,
    "args": {
      "External id": 2003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969772.770, "dur": 0.444,
    "args": {
      "External id": 2004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969777.524, "dur": 4.420,
    "args": {
      "External id": 2005,"Record function id": 0, "Ev Idx": 1492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969779.280, "dur": 1.981,
    "args": {
      "External id": 2006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969779.939, "dur": 0.937,
    "args": {
      "External id": 2007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969780.304, "dur": 0.449,
    "args": {
      "External id": 2008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969785.457, "dur": 4.663,
    "args": {
      "External id": 2009,"Record function id": 0, "Ev Idx": 1496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969787.263, "dur": 2.173,
    "args": {
      "External id": 2010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969787.953, "dur": 0.941,
    "args": {
      "External id": 2011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969788.329, "dur": 0.435,
    "args": {
      "External id": 2012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969793.077, "dur": 5.180,
    "args": {
      "External id": 2013,"Record function id": 0, "Ev Idx": 1500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969795.095, "dur": 2.426,
    "args": {
      "External id": 2014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969796.114, "dur": 0.925,
    "args": {
      "External id": 2015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969796.470, "dur": 0.445,
    "args": {
      "External id": 2016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969801.683, "dur": 4.990,
    "args": {
      "External id": 2017,"Record function id": 0, "Ev Idx": 1504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969803.611, "dur": 2.382,
    "args": {
      "External id": 2018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969804.464, "dur": 0.971,
    "args": {
      "External id": 2019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969804.844, "dur": 0.465,
    "args": {
      "External id": 2020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969809.916, "dur": 4.817,
    "args": {
      "External id": 2021,"Record function id": 0, "Ev Idx": 1508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969811.874, "dur": 2.034,
    "args": {
      "External id": 2022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969812.503, "dur": 0.877,
    "args": {
      "External id": 2023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969812.864, "dur": 0.372,
    "args": {
      "External id": 2024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969817.823, "dur": 4.417,
    "args": {
      "External id": 2025,"Record function id": 0, "Ev Idx": 1512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969819.557, "dur": 1.996,
    "args": {
      "External id": 2026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 1513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969820.330, "dur": 0.910,
    "args": {
      "External id": 2027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 1514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969820.688, "dur": 0.414,
    "args": {
      "External id": 2028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 1515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969825.452, "dur": 5.550,
    "args": {
      "External id": 2029,"Record function id": 0, "Ev Idx": 1516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969827.568, "dur": 2.559,
    "args": {
      "External id": 2030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969828.458, "dur": 0.965,
    "args": {
      "External id": 2031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969828.826, "dur": 0.453,
    "args": {
      "External id": 2032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969834.980, "dur": 5.539,
    "args": {
      "External id": 2033,"Record function id": 0, "Ev Idx": 1520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969837.160, "dur": 2.474,
    "args": {
      "External id": 2034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969838.148, "dur": 0.848,
    "args": {
      "External id": 2035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969838.538, "dur": 0.311,
    "args": {
      "External id": 2036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969843.786, "dur": 5.307,
    "args": {
      "External id": 2037,"Record function id": 0, "Ev Idx": 1524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969845.961, "dur": 2.287,
    "args": {
      "External id": 2038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969846.725, "dur": 0.975,
    "args": {
      "External id": 2039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969847.100, "dur": 0.475,
    "args": {
      "External id": 2040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969852.661, "dur": 4.905,
    "args": {
      "External id": 2041,"Record function id": 0, "Ev Idx": 1528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969854.501, "dur": 2.239,
    "args": {
      "External id": 2042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969855.250, "dur": 0.931,
    "args": {
      "External id": 2043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969855.622, "dur": 0.420,
    "args": {
      "External id": 2044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969860.777, "dur": 4.580,
    "args": {
      "External id": 2045,"Record function id": 0, "Ev Idx": 1532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507969862.521, "dur": 2.132,
    "args": {
      "External id": 2046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969863.233, "dur": 0.883,
    "args": {
      "External id": 2047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507969863.629, "dur": 0.345,
    "args": {
      "External id": 2048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507970146.478, "dur": 58.050,
    "args": {
      "External id": 2049,"Record function id": 0, "Ev Idx": 1536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507970187.252, "dur": 14.421,
    "args": {
      "External id": 2050,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507970193.114, "dur": 7.080,
    "args": {
      "External id": 2051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507970194.712, "dur": 3.589,
    "args": {
      "External id": 2052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507970225.691, "dur": 6.447,
    "args": {
      "External id": 2053,"Record function id": 0, "Ev Idx": 1540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507970228.013, "dur": 3.155,
    "args": {
      "External id": 2054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507970228.928, "dur": 1.592,
    "args": {
      "External id": 2055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507970229.544, "dur": 0.849,
    "args": {
      "External id": 2056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507970236.893, "dur": 264.317,
    "args": {
      "External id": 2057,"Record function id": 0, "Sequence number": 4686, "Fwd thread id": 1, "Ev Idx": 1544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507970240.017, "dur": 206.625,
    "args": {
      "External id": 2058,"Sequence number": 4686, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 1545
    }
  },
  {
    "ph": "f", "id": 4, "pid": 14565, "tid": 14640, "ts": 1290507970240.017,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290507970337.390, "dur": 93.397,
    "args": {
      "External id": 2059,"Record function id": 0, "scope": "user_scope", "Ev Idx": 1546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290507970463.944, "dur": 34.042,
    "args": {
      "External id": 2060,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 1547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507970507.399, "dur": 7.113,
    "args": {
      "External id": 2061,"Record function id": 0, "Ev Idx": 1548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507970509.820, "dur": 3.824,
    "args": {
      "External id": 2062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 1549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507970511.242, "dur": 2.038,
    "args": {
      "External id": 2063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 1550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507970512.038, "dur": 1.084,
    "args": {
      "External id": 2064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 1551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507974539.762, "dur": 2849.278,
    "args": {
      "External id": 2065,"Record function id": 0, "Sequence number": 4689, "Fwd thread id": 1, "Ev Idx": 1552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507974546.515, "dur": 2668.586,
    "args": {
      "External id": 2066,"Sequence number": 4689, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 1553
    }
  },
  {
    "ph": "f", "id": 5, "pid": 14565, "tid": 14640, "ts": 1290507974546.515,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290507975000.428, "dur": 2173.750,
    "args": {
      "External id": 2067,"Record function id": 0, "scope": "user_scope", "Ev Idx": 1554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290507975862.173, "dur": 23.107,
    "args": {
      "External id": 2068,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 1555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977417.315, "dur": 17.217,
    "args": {
      "External id": 2069,"Record function id": 0, "Ev Idx": 1556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977422.675, "dur": 9.781,
    "args": {
      "External id": 2070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977426.767, "dur": 4.613,
    "args": {
      "External id": 2071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977428.539, "dur": 2.539,
    "args": {
      "External id": 2072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977440.709, "dur": 9.473,
    "args": {
      "External id": 2073,"Record function id": 0, "Ev Idx": 1560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977443.211, "dur": 5.403,
    "args": {
      "External id": 2074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977444.873, "dur": 3.170,
    "args": {
      "External id": 2075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977447.203, "dur": 0.600,
    "args": {
      "External id": 2076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977455.353, "dur": 9.307,
    "args": {
      "External id": 2077,"Record function id": 0, "Ev Idx": 1564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977458.138, "dur": 5.214,
    "args": {
      "External id": 2078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977459.583, "dur": 3.069,
    "args": {
      "External id": 2079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977461.643, "dur": 0.724,
    "args": {
      "External id": 2080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977469.825, "dur": 6.905,
    "args": {
      "External id": 2081,"Record function id": 0, "Ev Idx": 1568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977471.895, "dur": 3.675,
    "args": {
      "External id": 2082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977473.447, "dur": 1.528,
    "args": {
      "External id": 2083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977474.127, "dur": 0.549,
    "args": {
      "External id": 2084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977481.745, "dur": 7.843,
    "args": {
      "External id": 2085,"Record function id": 0, "Ev Idx": 1572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977484.040, "dur": 4.653,
    "args": {
      "External id": 2086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977485.302, "dur": 2.888,
    "args": {
      "External id": 2087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977486.321, "dur": 0.635,
    "args": {
      "External id": 2088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977493.179, "dur": 4.810,
    "args": {
      "External id": 2089,"Record function id": 0, "Ev Idx": 1576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977494.897, "dur": 2.384,
    "args": {
      "External id": 2090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977495.653, "dur": 1.275,
    "args": {
      "External id": 2091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977496.233, "dur": 0.549,
    "args": {
      "External id": 2092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977501.232, "dur": 4.653,
    "args": {
      "External id": 2093,"Record function id": 0, "Ev Idx": 1580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977502.932, "dur": 2.253,
    "args": {
      "External id": 2094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977503.738, "dur": 0.911,
    "args": {
      "External id": 2095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977504.103, "dur": 0.392,
    "args": {
      "External id": 2096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977510.568, "dur": 7.551,
    "args": {
      "External id": 2097,"Record function id": 0, "Ev Idx": 1584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977513.017, "dur": 3.894,
    "args": {
      "External id": 2098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977514.213, "dur": 2.010,
    "args": {
      "External id": 2099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977514.872, "dur": 1.056,
    "args": {
      "External id": 2100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977523.198, "dur": 8.663,
    "args": {
      "External id": 2101,"Record function id": 0, "Ev Idx": 1588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977526.879, "dur": 3.789,
    "args": {
      "External id": 2102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977528.244, "dur": 1.861,
    "args": {
      "External id": 2103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977529.091, "dur": 0.758,
    "args": {
      "External id": 2104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977537.194, "dur": 8.136,
    "args": {
      "External id": 2105,"Record function id": 0, "Ev Idx": 1592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977539.532, "dur": 4.504,
    "args": {
      "External id": 2106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977541.380, "dur": 1.391,
    "args": {
      "External id": 2107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977542.017, "dur": 0.504,
    "args": {
      "External id": 2108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977550.743, "dur": 6.847,
    "args": {
      "External id": 2109,"Record function id": 0, "Ev Idx": 1596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977552.704, "dur": 3.655,
    "args": {
      "External id": 2110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977554.215, "dur": 1.489,
    "args": {
      "External id": 2111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977554.965, "dur": 0.487,
    "args": {
      "External id": 2112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977563.554, "dur": 6.483,
    "args": {
      "External id": 2113,"Record function id": 0, "Ev Idx": 1600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977565.681, "dur": 3.267,
    "args": {
      "External id": 2114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977566.882, "dur": 1.526,
    "args": {
      "External id": 2115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977567.493, "dur": 0.648,
    "args": {
      "External id": 2116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977575.741, "dur": 105.452,
    "args": {
      "External id": 2117,"Record function id": 0, "Ev Idx": 1604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977578.347, "dur": 93.689,
    "args": {
      "External id": 2118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977579.780, "dur": 88.714,
    "args": {
      "External id": 2119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977580.593, "dur": 79.760,
    "args": {
      "External id": 2120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977708.748, "dur": 16.226,
    "args": {
      "External id": 2121,"Record function id": 0, "Ev Idx": 1608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977716.787, "dur": 7.314,
    "args": {
      "External id": 2122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977720.339, "dur": 3.171,
    "args": {
      "External id": 2123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977721.826, "dur": 1.505,
    "args": {
      "External id": 2124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977729.593, "dur": 5.899,
    "args": {
      "External id": 2125,"Record function id": 0, "Ev Idx": 1612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977731.901, "dur": 2.800,
    "args": {
      "External id": 2126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977733.051, "dur": 1.088,
    "args": {
      "External id": 2127,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977733.702, "dur": 0.299,
    "args": {
      "External id": 2128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977742.296, "dur": 6.281,
    "args": {
      "External id": 2129,"Record function id": 0, "Ev Idx": 1616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977744.588, "dur": 3.097,
    "args": {
      "External id": 2130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977745.624, "dur": 1.521,
    "args": {
      "External id": 2131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977746.490, "dur": 0.524,
    "args": {
      "External id": 2132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977752.209, "dur": 5.334,
    "args": {
      "External id": 2133,"Record function id": 0, "Ev Idx": 1620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977753.926, "dur": 2.805,
    "args": {
      "External id": 2134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977754.759, "dur": 1.492,
    "args": {
      "External id": 2135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977755.576, "dur": 0.480,
    "args": {
      "External id": 2136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977761.125, "dur": 4.614,
    "args": {
      "External id": 2137,"Record function id": 0, "Ev Idx": 1624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977762.659, "dur": 2.344,
    "args": {
      "External id": 2138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977763.536, "dur": 0.926,
    "args": {
      "External id": 2139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977763.923, "dur": 0.412,
    "args": {
      "External id": 2140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977769.354, "dur": 4.943,
    "args": {
      "External id": 2141,"Record function id": 0, "Ev Idx": 1628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977771.082, "dur": 2.370,
    "args": {
      "External id": 2142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977771.972, "dur": 0.904,
    "args": {
      "External id": 2143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977772.434, "dur": 0.309,
    "args": {
      "External id": 2144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977777.717, "dur": 4.229,
    "args": {
      "External id": 2145,"Record function id": 0, "Ev Idx": 1632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977779.163, "dur": 1.967,
    "args": {
      "External id": 2146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977779.876, "dur": 0.775,
    "args": {
      "External id": 2147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977780.224, "dur": 0.300,
    "args": {
      "External id": 2148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977785.264, "dur": 5.996,
    "args": {
      "External id": 2149,"Record function id": 0, "Ev Idx": 1636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977787.059, "dur": 3.391,
    "args": {
      "External id": 2150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977788.073, "dur": 0.943,
    "args": {
      "External id": 2151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977788.583, "dur": 0.302,
    "args": {
      "External id": 2152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977794.495, "dur": 4.289,
    "args": {
      "External id": 2153,"Record function id": 0, "Ev Idx": 1640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977796.005, "dur": 1.932,
    "args": {
      "External id": 2154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977796.777, "dur": 0.750,
    "args": {
      "External id": 2155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977797.106, "dur": 0.296,
    "args": {
      "External id": 2156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977801.937, "dur": 4.198,
    "args": {
      "External id": 2157,"Record function id": 0, "Ev Idx": 1644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977803.325, "dur": 2.070,
    "args": {
      "External id": 2158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977803.993, "dur": 0.930,
    "args": {
      "External id": 2159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977804.376, "dur": 0.421,
    "args": {
      "External id": 2160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977809.414, "dur": 4.604,
    "args": {
      "External id": 2161,"Record function id": 0, "Ev Idx": 1648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977810.992, "dur": 2.158,
    "args": {
      "External id": 2162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977811.822, "dur": 0.797,
    "args": {
      "External id": 2163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977812.191, "dur": 0.300,
    "args": {
      "External id": 2164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977817.223, "dur": 4.477,
    "args": {
      "External id": 2165,"Record function id": 0, "Ev Idx": 1652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977818.855, "dur": 2.058,
    "args": {
      "External id": 2166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977819.595, "dur": 0.810,
    "args": {
      "External id": 2167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977819.968, "dur": 0.306,
    "args": {
      "External id": 2168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977824.864, "dur": 7.537,
    "args": {
      "External id": 2169,"Record function id": 0, "Ev Idx": 1656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977827.511, "dur": 3.998,
    "args": {
      "External id": 2170,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977828.235, "dur": 2.775,
    "args": {
      "External id": 2171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977828.739, "dur": 2.100,
    "args": {
      "External id": 2172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977835.725, "dur": 4.773,
    "args": {
      "External id": 2173,"Record function id": 0, "Ev Idx": 1660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977837.436, "dur": 2.338,
    "args": {
      "External id": 2174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977838.171, "dur": 1.042,
    "args": {
      "External id": 2175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977838.688, "dur": 0.388,
    "args": {
      "External id": 2176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977843.940, "dur": 4.678,
    "args": {
      "External id": 2177,"Record function id": 0, "Ev Idx": 1664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977845.573, "dur": 2.200,
    "args": {
      "External id": 2178,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977846.198, "dur": 1.035,
    "args": {
      "External id": 2179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977846.586, "dur": 0.502,
    "args": {
      "External id": 2180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977851.810, "dur": 4.938,
    "args": {
      "External id": 2181,"Record function id": 0, "Ev Idx": 1668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977853.599, "dur": 2.151,
    "args": {
      "External id": 2182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977854.312, "dur": 0.977,
    "args": {
      "External id": 2183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977854.814, "dur": 0.345,
    "args": {
      "External id": 2184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977860.134, "dur": 5.572,
    "args": {
      "External id": 2185,"Record function id": 0, "Ev Idx": 1672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977861.843, "dur": 3.040,
    "args": {
      "External id": 2186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977862.802, "dur": 1.567,
    "args": {
      "External id": 2187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977863.197, "dur": 1.040,
    "args": {
      "External id": 2188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977868.990, "dur": 4.742,
    "args": {
      "External id": 2189,"Record function id": 0, "Ev Idx": 1676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977870.450, "dur": 2.465,
    "args": {
      "External id": 2190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977871.180, "dur": 1.300,
    "args": {
      "External id": 2191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977871.555, "dur": 0.785,
    "args": {
      "External id": 2192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977877.107, "dur": 4.976,
    "args": {
      "External id": 2193,"Record function id": 0, "Ev Idx": 1680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977879.040, "dur": 2.259,
    "args": {
      "External id": 2194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977879.890, "dur": 0.833,
    "args": {
      "External id": 2195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977880.275, "dur": 0.322,
    "args": {
      "External id": 2196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977885.851, "dur": 4.854,
    "args": {
      "External id": 2197,"Record function id": 0, "Ev Idx": 1684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977887.616, "dur": 2.317,
    "args": {
      "External id": 2198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977888.484, "dur": 0.803,
    "args": {
      "External id": 2199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977888.842, "dur": 0.306,
    "args": {
      "External id": 2200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977894.339, "dur": 88.879,
    "args": {
      "External id": 2201,"Record function id": 0, "Ev Idx": 1688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507977895.798, "dur": 77.566,
    "args": {
      "External id": 2202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977896.541, "dur": 73.744,
    "args": {
      "External id": 2203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507977896.918, "dur": 68.929,
    "args": {
      "External id": 2204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978013.645, "dur": 17.170,
    "args": {
      "External id": 2205,"Record function id": 0, "Ev Idx": 1692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978022.077, "dur": 7.598,
    "args": {
      "External id": 2206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978026.122, "dur": 2.946,
    "args": {
      "External id": 2207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978027.532, "dur": 1.379,
    "args": {
      "External id": 2208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978035.568, "dur": 7.279,
    "args": {
      "External id": 2209,"Record function id": 0, "Ev Idx": 1696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978038.597, "dur": 3.463,
    "args": {
      "External id": 2210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978039.821, "dur": 1.770,
    "args": {
      "External id": 2211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978040.785, "dur": 0.664,
    "args": {
      "External id": 2212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978046.754, "dur": 5.508,
    "args": {
      "External id": 2213,"Record function id": 0, "Ev Idx": 1700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978048.583, "dur": 2.803,
    "args": {
      "External id": 2214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978049.598, "dur": 1.286,
    "args": {
      "External id": 2215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978050.431, "dur": 0.323,
    "args": {
      "External id": 2216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978056.022, "dur": 4.351,
    "args": {
      "External id": 2217,"Record function id": 0, "Ev Idx": 1704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978057.538, "dur": 2.073,
    "args": {
      "External id": 2218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978058.258, "dur": 0.901,
    "args": {
      "External id": 2219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978058.638, "dur": 0.367,
    "args": {
      "External id": 2220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978063.837, "dur": 4.265,
    "args": {
      "External id": 2221,"Record function id": 0, "Ev Idx": 1708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978065.509, "dur": 1.895,
    "args": {
      "External id": 2222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978066.164, "dur": 0.814,
    "args": {
      "External id": 2223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978066.523, "dur": 0.327,
    "args": {
      "External id": 2224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978071.532, "dur": 4.835,
    "args": {
      "External id": 2225,"Record function id": 0, "Ev Idx": 1712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978073.240, "dur": 2.409,
    "args": {
      "External id": 2226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978074.094, "dur": 1.083,
    "args": {
      "External id": 2227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978074.587, "dur": 0.439,
    "args": {
      "External id": 2228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978083.602, "dur": 4.445,
    "args": {
      "External id": 2229,"Record function id": 0, "Ev Idx": 1716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978085.148, "dur": 2.070,
    "args": {
      "External id": 2230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978085.878, "dur": 0.912,
    "args": {
      "External id": 2231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978086.284, "dur": 0.378,
    "args": {
      "External id": 2232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978091.434, "dur": 4.638,
    "args": {
      "External id": 2233,"Record function id": 0, "Ev Idx": 1720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978093.112, "dur": 2.189,
    "args": {
      "External id": 2234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978094.011, "dur": 0.849,
    "args": {
      "External id": 2235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978094.396, "dur": 0.323,
    "args": {
      "External id": 2236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978100.434, "dur": 6.233,
    "args": {
      "External id": 2237,"Record function id": 0, "Ev Idx": 1724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978102.541, "dur": 3.356,
    "args": {
      "External id": 2238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978103.762, "dur": 1.067,
    "args": {
      "External id": 2239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978104.347, "dur": 0.303,
    "args": {
      "External id": 2240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978110.025, "dur": 4.697,
    "args": {
      "External id": 2241,"Record function id": 0, "Ev Idx": 1728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978111.671, "dur": 2.307,
    "args": {
      "External id": 2242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978112.554, "dur": 0.869,
    "args": {
      "External id": 2243,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978112.928, "dur": 0.313,
    "args": {
      "External id": 2244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978117.950, "dur": 116.636,
    "args": {
      "External id": 2245,"Record function id": 0, "Ev Idx": 1732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978119.406, "dur": 107.564,
    "args": {
      "External id": 2246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978120.189, "dur": 104.044,
    "args": {
      "External id": 2247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978216.356, "dur": 6.545,
    "args": {
      "External id": 2248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978254.463, "dur": 11.772,
    "args": {
      "External id": 2249,"Record function id": 0, "Ev Idx": 1736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978259.227, "dur": 6.004,
    "args": {
      "External id": 2250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978262.277, "dur": 2.322,
    "args": {
      "External id": 2251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978263.363, "dur": 1.051,
    "args": {
      "External id": 2252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978270.871, "dur": 6.289,
    "args": {
      "External id": 2253,"Record function id": 0, "Ev Idx": 1740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978273.009, "dur": 3.342,
    "args": {
      "External id": 2254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978274.249, "dur": 1.654,
    "args": {
      "External id": 2255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978274.923, "dur": 0.835,
    "args": {
      "External id": 2256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978280.958, "dur": 5.377,
    "args": {
      "External id": 2257,"Record function id": 0, "Ev Idx": 1744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978283.075, "dur": 2.569,
    "args": {
      "External id": 2258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978283.914, "dur": 1.241,
    "args": {
      "External id": 2259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978284.459, "dur": 0.566,
    "args": {
      "External id": 2260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978290.098, "dur": 6.413,
    "args": {
      "External id": 2261,"Record function id": 0, "Ev Idx": 1748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978292.071, "dur": 3.700,
    "args": {
      "External id": 2262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978292.857, "dur": 2.365,
    "args": {
      "External id": 2263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978294.531, "dur": 0.562,
    "args": {
      "External id": 2264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978300.397, "dur": 6.806,
    "args": {
      "External id": 2265,"Record function id": 0, "Ev Idx": 1752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978302.318, "dur": 4.034,
    "args": {
      "External id": 2266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978303.283, "dur": 2.559,
    "args": {
      "External id": 2267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978303.821, "dur": 1.865,
    "args": {
      "External id": 2268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978311.109, "dur": 4.893,
    "args": {
      "External id": 2269,"Record function id": 0, "Ev Idx": 1756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978313.138, "dur": 2.179,
    "args": {
      "External id": 2270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978313.893, "dur": 0.912,
    "args": {
      "External id": 2271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978314.296, "dur": 0.348,
    "args": {
      "External id": 2272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978319.819, "dur": 5.036,
    "args": {
      "External id": 2273,"Record function id": 0, "Ev Idx": 1760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978321.423, "dur": 2.668,
    "args": {
      "External id": 2274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978322.186, "dur": 1.332,
    "args": {
      "External id": 2275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978322.579, "dur": 0.809,
    "args": {
      "External id": 2276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978328.562, "dur": 228.133,
    "args": {
      "External id": 2277,"Record function id": 0, "Ev Idx": 1764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978330.038, "dur": 214.539,
    "args": {
      "External id": 2278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978436.727, "dur": 103.762,
    "args": {
      "External id": 2279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978440.813, "dur": 95.419,
    "args": {
      "External id": 2280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978581.589, "dur": 17.931,
    "args": {
      "External id": 2281,"Record function id": 0, "Ev Idx": 1768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978589.982, "dur": 8.665,
    "args": {
      "External id": 2282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978594.152, "dur": 2.968,
    "args": {
      "External id": 2283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978595.891, "dur": 1.058,
    "args": {
      "External id": 2284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978604.381, "dur": 6.429,
    "args": {
      "External id": 2285,"Record function id": 0, "Ev Idx": 1772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978607.036, "dur": 3.014,
    "args": {
      "External id": 2286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978608.116, "dur": 1.230,
    "args": {
      "External id": 2287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978608.840, "dur": 0.371,
    "args": {
      "External id": 2288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978614.826, "dur": 7.016,
    "args": {
      "External id": 2289,"Record function id": 0, "Ev Idx": 1776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978616.876, "dur": 4.082,
    "args": {
      "External id": 2290,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978618.757, "dur": 1.453,
    "args": {
      "External id": 2291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978619.605, "dur": 0.473,
    "args": {
      "External id": 2292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978625.708, "dur": 5.488,
    "args": {
      "External id": 2293,"Record function id": 0, "Ev Idx": 1780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978627.536, "dur": 2.818,
    "args": {
      "External id": 2294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978628.844, "dur": 1.074,
    "args": {
      "External id": 2295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978629.448, "dur": 0.330,
    "args": {
      "External id": 2296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978634.758, "dur": 4.604,
    "args": {
      "External id": 2297,"Record function id": 0, "Ev Idx": 1784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978636.252, "dur": 2.303,
    "args": {
      "External id": 2298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978637.210, "dur": 0.909,
    "args": {
      "External id": 2299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978637.645, "dur": 0.332,
    "args": {
      "External id": 2300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978642.890, "dur": 4.799,
    "args": {
      "External id": 2301,"Record function id": 0, "Ev Idx": 1788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978644.547, "dur": 2.421,
    "args": {
      "External id": 2302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978645.522, "dur": 0.945,
    "args": {
      "External id": 2303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978645.912, "dur": 0.426,
    "args": {
      "External id": 2304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978651.367, "dur": 4.851,
    "args": {
      "External id": 2305,"Record function id": 0, "Ev Idx": 1792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978653.299, "dur": 2.050,
    "args": {
      "External id": 2306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978654.045, "dur": 0.869,
    "args": {
      "External id": 2307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978654.433, "dur": 0.343,
    "args": {
      "External id": 2308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978659.826, "dur": 4.794,
    "args": {
      "External id": 2309,"Record function id": 0, "Ev Idx": 1796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978661.498, "dur": 2.387,
    "args": {
      "External id": 2310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978662.517, "dur": 0.866,
    "args": {
      "External id": 2311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978662.892, "dur": 0.366,
    "args": {
      "External id": 2312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978668.294, "dur": 4.773,
    "args": {
      "External id": 2313,"Record function id": 0, "Ev Idx": 1800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978669.958, "dur": 2.223,
    "args": {
      "External id": 2314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978670.860, "dur": 0.832,
    "args": {
      "External id": 2315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978671.197, "dur": 0.350,
    "args": {
      "External id": 2316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978676.481, "dur": 5.548,
    "args": {
      "External id": 2317,"Record function id": 0, "Ev Idx": 1804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978679.089, "dur": 2.122,
    "args": {
      "External id": 2318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978679.858, "dur": 0.917,
    "args": {
      "External id": 2319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978680.218, "dur": 0.416,
    "args": {
      "External id": 2320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978685.523, "dur": 4.476,
    "args": {
      "External id": 2321,"Record function id": 0, "Ev Idx": 1808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978686.952, "dur": 2.254,
    "args": {
      "External id": 2322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978687.851, "dur": 0.903,
    "args": {
      "External id": 2323,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978688.272, "dur": 0.322,
    "args": {
      "External id": 2324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978693.416, "dur": 5.161,
    "args": {
      "External id": 2325,"Record function id": 0, "Ev Idx": 1812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978695.520, "dur": 2.306,
    "args": {
      "External id": 2326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978696.334, "dur": 0.840,
    "args": {
      "External id": 2327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978696.730, "dur": 0.312,
    "args": {
      "External id": 2328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978702.140, "dur": 4.921,
    "args": {
      "External id": 2329,"Record function id": 0, "Ev Idx": 1816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978703.780, "dur": 2.385,
    "args": {
      "External id": 2330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978704.690, "dur": 0.819,
    "args": {
      "External id": 2331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978705.074, "dur": 0.303,
    "args": {
      "External id": 2332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978711.398, "dur": 4.875,
    "args": {
      "External id": 2333,"Record function id": 0, "Ev Idx": 1820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978712.880, "dur": 2.603,
    "args": {
      "External id": 2334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978713.499, "dur": 1.502,
    "args": {
      "External id": 2335,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978713.916, "dur": 0.912,
    "args": {
      "External id": 2336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978720.065, "dur": 4.810,
    "args": {
      "External id": 2337,"Record function id": 0, "Ev Idx": 1824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978721.818, "dur": 2.199,
    "args": {
      "External id": 2338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978722.570, "dur": 0.970,
    "args": {
      "External id": 2339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978722.978, "dur": 0.422,
    "args": {
      "External id": 2340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978728.441, "dur": 4.772,
    "args": {
      "External id": 2341,"Record function id": 0, "Ev Idx": 1828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978729.987, "dur": 2.358,
    "args": {
      "External id": 2342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978730.887, "dur": 0.872,
    "args": {
      "External id": 2343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978731.299, "dur": 0.330,
    "args": {
      "External id": 2344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978739.267, "dur": 4.888,
    "args": {
      "External id": 2345,"Record function id": 0, "Ev Idx": 1832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978740.799, "dur": 2.475,
    "args": {
      "External id": 2346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978741.704, "dur": 1.054,
    "args": {
      "External id": 2347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978742.274, "dur": 0.348,
    "args": {
      "External id": 2348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978747.653, "dur": 5.534,
    "args": {
      "External id": 2349,"Record function id": 0, "Ev Idx": 1836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978749.371, "dur": 2.851,
    "args": {
      "External id": 2350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978750.165, "dur": 1.459,
    "args": {
      "External id": 2351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978750.559, "dur": 0.926,
    "args": {
      "External id": 2352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978756.534, "dur": 5.572,
    "args": {
      "External id": 2353,"Record function id": 0, "Ev Idx": 1840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978758.262, "dur": 3.018,
    "args": {
      "External id": 2354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978758.987, "dur": 1.829,
    "args": {
      "External id": 2355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978759.539, "dur": 1.122,
    "args": {
      "External id": 2356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978765.586, "dur": 5.353,
    "args": {
      "External id": 2357,"Record function id": 0, "Ev Idx": 1844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978767.580, "dur": 2.608,
    "args": {
      "External id": 2358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978768.674, "dur": 0.808,
    "args": {
      "External id": 2359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978769.018, "dur": 0.332,
    "args": {
      "External id": 2360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978774.755, "dur": 80.945,
    "args": {
      "External id": 2361,"Record function id": 0, "Ev Idx": 1848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978776.967, "dur": 71.637,
    "args": {
      "External id": 2362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978777.778, "dur": 66.111,
    "args": {
      "External id": 2363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978778.114, "dur": 61.723,
    "args": {
      "External id": 2364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978880.461, "dur": 16.306,
    "args": {
      "External id": 2365,"Record function id": 0, "Ev Idx": 1852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978888.912, "dur": 6.872,
    "args": {
      "External id": 2366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978892.412, "dur": 2.722,
    "args": {
      "External id": 2367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978893.869, "dur": 1.118,
    "args": {
      "External id": 2368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978901.621, "dur": 9.202,
    "args": {
      "External id": 2369,"Record function id": 0, "Ev Idx": 1856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978904.098, "dur": 5.925,
    "args": {
      "External id": 2370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978905.265, "dur": 4.056,
    "args": {
      "External id": 2371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978908.498, "dur": 0.682,
    "args": {
      "External id": 2372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978914.929, "dur": 5.819,
    "args": {
      "External id": 2373,"Record function id": 0, "Ev Idx": 1860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978917.075, "dur": 2.811,
    "args": {
      "External id": 2374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978918.044, "dur": 1.203,
    "args": {
      "External id": 2375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978918.572, "dur": 0.525,
    "args": {
      "External id": 2376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978924.436, "dur": 5.894,
    "args": {
      "External id": 2377,"Record function id": 0, "Ev Idx": 1864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978926.709, "dur": 2.851,
    "args": {
      "External id": 2378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978927.668, "dur": 1.386,
    "args": {
      "External id": 2379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978928.440, "dur": 0.476,
    "args": {
      "External id": 2380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978933.753, "dur": 4.580,
    "args": {
      "External id": 2381,"Record function id": 0, "Ev Idx": 1868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978935.266, "dur": 2.345,
    "args": {
      "External id": 2382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978936.080, "dur": 0.970,
    "args": {
      "External id": 2383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978936.625, "dur": 0.292,
    "args": {
      "External id": 2384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978941.782, "dur": 4.501,
    "args": {
      "External id": 2385,"Record function id": 0, "Ev Idx": 1872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978943.327, "dur": 2.223,
    "args": {
      "External id": 2386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978944.107, "dur": 0.957,
    "args": {
      "External id": 2387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978944.611, "dur": 0.319,
    "args": {
      "External id": 2388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978949.584, "dur": 4.378,
    "args": {
      "External id": 2389,"Record function id": 0, "Ev Idx": 1876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978951.079, "dur": 2.125,
    "args": {
      "External id": 2390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978951.855, "dur": 0.820,
    "args": {
      "External id": 2391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978952.229, "dur": 0.315,
    "args": {
      "External id": 2392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978957.307, "dur": 4.303,
    "args": {
      "External id": 2393,"Record function id": 0, "Ev Idx": 1880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978958.670, "dur": 2.094,
    "args": {
      "External id": 2394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978959.375, "dur": 0.864,
    "args": {
      "External id": 2395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978959.732, "dur": 0.384,
    "args": {
      "External id": 2396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978965.028, "dur": 4.338,
    "args": {
      "External id": 2397,"Record function id": 0, "Ev Idx": 1884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978966.453, "dur": 2.050,
    "args": {
      "External id": 2398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978967.068, "dur": 0.894,
    "args": {
      "External id": 2399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978967.436, "dur": 0.391,
    "args": {
      "External id": 2400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978972.765, "dur": 5.222,
    "args": {
      "External id": 2401,"Record function id": 0, "Ev Idx": 1888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978974.596, "dur": 2.653,
    "args": {
      "External id": 2402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978975.524, "dur": 1.190,
    "args": {
      "External id": 2403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978976.010, "dur": 0.578,
    "args": {
      "External id": 2404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978981.439, "dur": 4.582,
    "args": {
      "External id": 2405,"Record function id": 0, "Ev Idx": 1892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978982.974, "dur": 2.319,
    "args": {
      "External id": 2406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978983.898, "dur": 0.873,
    "args": {
      "External id": 2407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978984.281, "dur": 0.340,
    "args": {
      "External id": 2408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978989.506, "dur": 5.651,
    "args": {
      "External id": 2409,"Record function id": 0, "Ev Idx": 1896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978991.653, "dur": 2.604,
    "args": {
      "External id": 2410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978992.713, "dur": 1.015,
    "args": {
      "External id": 2411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507978993.294, "dur": 0.302,
    "args": {
      "External id": 2412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507978998.546, "dur": 5.043,
    "args": {
      "External id": 2413,"Record function id": 0, "Ev Idx": 1900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979000.192, "dur": 2.528,
    "args": {
      "External id": 2414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979001.190, "dur": 0.887,
    "args": {
      "External id": 2415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979001.647, "dur": 0.305,
    "args": {
      "External id": 2416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979008.022, "dur": 5.301,
    "args": {
      "External id": 2417,"Record function id": 0, "Ev Idx": 1904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979009.799, "dur": 2.684,
    "args": {
      "External id": 2418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979010.519, "dur": 1.379,
    "args": {
      "External id": 2419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979010.891, "dur": 0.830,
    "args": {
      "External id": 2420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979016.850, "dur": 5.011,
    "args": {
      "External id": 2421,"Record function id": 0, "Ev Idx": 1908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979018.453, "dur": 2.695,
    "args": {
      "External id": 2422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979019.398, "dur": 1.250,
    "args": {
      "External id": 2423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979020.094, "dur": 0.421,
    "args": {
      "External id": 2424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979025.497, "dur": 4.476,
    "args": {
      "External id": 2425,"Record function id": 0, "Ev Idx": 1912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979027.070, "dur": 2.121,
    "args": {
      "External id": 2426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979027.847, "dur": 0.835,
    "args": {
      "External id": 2427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979028.222, "dur": 0.327,
    "args": {
      "External id": 2428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979033.549, "dur": 4.634,
    "args": {
      "External id": 2429,"Record function id": 0, "Ev Idx": 1916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979035.353, "dur": 2.052,
    "args": {
      "External id": 2430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979036.124, "dur": 0.797,
    "args": {
      "External id": 2431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979036.492, "dur": 0.307,
    "args": {
      "External id": 2432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979041.873, "dur": 4.694,
    "args": {
      "External id": 2433,"Record function id": 0, "Ev Idx": 1920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979043.761, "dur": 2.019,
    "args": {
      "External id": 2434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979044.453, "dur": 0.840,
    "args": {
      "External id": 2435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979044.830, "dur": 0.319,
    "args": {
      "External id": 2436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979050.760, "dur": 4.442,
    "args": {
      "External id": 2437,"Record function id": 0, "Ev Idx": 1924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979052.263, "dur": 2.099,
    "args": {
      "External id": 2438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979052.971, "dur": 0.864,
    "args": {
      "External id": 2439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979053.345, "dur": 0.366,
    "args": {
      "External id": 2440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979058.973, "dur": 4.723,
    "args": {
      "External id": 2441,"Record function id": 0, "Ev Idx": 1928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979060.684, "dur": 2.252,
    "args": {
      "External id": 2442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979061.544, "dur": 0.818,
    "args": {
      "External id": 2443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979061.924, "dur": 0.310,
    "args": {
      "External id": 2444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979067.620, "dur": 4.236,
    "args": {
      "External id": 2445,"Record function id": 0, "Ev Idx": 1932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979069.071, "dur": 1.985,
    "args": {
      "External id": 2446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979069.782, "dur": 0.799,
    "args": {
      "External id": 2447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979070.171, "dur": 0.270,
    "args": {
      "External id": 2448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979075.260, "dur": 47.103,
    "args": {
      "External id": 2449,"Record function id": 0, "Ev Idx": 1936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979076.689, "dur": 44.257,
    "args": {
      "External id": 2450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979077.536, "dur": 41.105,
    "args": {
      "External id": 2451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979077.908, "dur": 39.463,
    "args": {
      "External id": 2452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 1939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979132.887, "dur": 7.772,
    "args": {
      "External id": 2453,"Record function id": 0, "Ev Idx": 1940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979137.244, "dur": 2.681,
    "args": {
      "External id": 2454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979138.145, "dur": 1.203,
    "args": {
      "External id": 2455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979138.696, "dur": 0.521,
    "args": {
      "External id": 2456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979144.474, "dur": 5.283,
    "args": {
      "External id": 2457,"Record function id": 0, "Ev Idx": 1944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979146.224, "dur": 2.817,
    "args": {
      "External id": 2458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979147.206, "dur": 1.091,
    "args": {
      "External id": 2459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979147.768, "dur": 0.356,
    "args": {
      "External id": 2460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979153.079, "dur": 4.703,
    "args": {
      "External id": 2461,"Record function id": 0, "Ev Idx": 1948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979154.898, "dur": 2.142,
    "args": {
      "External id": 2462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979155.650, "dur": 0.856,
    "args": {
      "External id": 2463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979156.004, "dur": 0.321,
    "args": {
      "External id": 2464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979160.964, "dur": 4.274,
    "args": {
      "External id": 2465,"Record function id": 0, "Ev Idx": 1952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979162.252, "dur": 2.294,
    "args": {
      "External id": 2466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979163.020, "dur": 0.983,
    "args": {
      "External id": 2467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979163.354, "dur": 0.504,
    "args": {
      "External id": 2468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979168.654, "dur": 4.402,
    "args": {
      "External id": 2469,"Record function id": 0, "Ev Idx": 1956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979170.031, "dur": 2.308,
    "args": {
      "External id": 2470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979170.764, "dur": 0.955,
    "args": {
      "External id": 2471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979171.129, "dur": 0.467,
    "args": {
      "External id": 2472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 1959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979176.521, "dur": 4.652,
    "args": {
      "External id": 2473,"Record function id": 0, "Ev Idx": 1960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979178.266, "dur": 2.202,
    "args": {
      "External id": 2474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 1961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979179.032, "dur": 0.888,
    "args": {
      "External id": 2475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 1962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979179.407, "dur": 0.391,
    "args": {
      "External id": 2476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 1963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979184.905, "dur": 4.494,
    "args": {
      "External id": 2477,"Record function id": 0, "Ev Idx": 1964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979186.552, "dur": 2.138,
    "args": {
      "External id": 2478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979187.228, "dur": 0.988,
    "args": {
      "External id": 2479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979187.769, "dur": 0.314,
    "args": {
      "External id": 2480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 1967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979193.049, "dur": 5.634,
    "args": {
      "External id": 2481,"Record function id": 0, "Ev Idx": 1968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979194.922, "dur": 2.965,
    "args": {
      "External id": 2482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979196.039, "dur": 1.171,
    "args": {
      "External id": 2483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979196.655, "dur": 0.428,
    "args": {
      "External id": 2484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979202.026, "dur": 4.980,
    "args": {
      "External id": 2485,"Record function id": 0, "Ev Idx": 1972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979203.826, "dur": 2.400,
    "args": {
      "External id": 2486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979204.873, "dur": 0.842,
    "args": {
      "External id": 2487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979205.233, "dur": 0.312,
    "args": {
      "External id": 2488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 1975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979210.460, "dur": 5.041,
    "args": {
      "External id": 2489,"Record function id": 0, "Ev Idx": 1976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979212.327, "dur": 2.378,
    "args": {
      "External id": 2490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979213.227, "dur": 0.995,
    "args": {
      "External id": 2491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979213.600, "dur": 0.490,
    "args": {
      "External id": 2492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 1979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979218.851, "dur": 4.665,
    "args": {
      "External id": 2493,"Record function id": 0, "Ev Idx": 1980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979220.372, "dur": 2.345,
    "args": {
      "External id": 2494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979221.273, "dur": 0.984,
    "args": {
      "External id": 2495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979221.664, "dur": 0.453,
    "args": {
      "External id": 2496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 1983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979226.968, "dur": 4.943,
    "args": {
      "External id": 2497,"Record function id": 0, "Ev Idx": 1984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979228.927, "dur": 2.264,
    "args": {
      "External id": 2498,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979229.719, "dur": 0.898,
    "args": {
      "External id": 2499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979230.100, "dur": 0.377,
    "args": {
      "External id": 2500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 1987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979235.404, "dur": 5.599,
    "args": {
      "External id": 2501,"Record function id": 0, "Ev Idx": 1988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979237.074, "dur": 3.194,
    "args": {
      "External id": 2502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979237.878, "dur": 1.850,
    "args": {
      "External id": 2503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979238.647, "dur": 0.937,
    "args": {
      "External id": 2504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 1991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979244.570, "dur": 5.247,
    "args": {
      "External id": 2505,"Record function id": 0, "Ev Idx": 1992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979246.465, "dur": 2.616,
    "args": {
      "External id": 2506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979247.364, "dur": 1.170,
    "args": {
      "External id": 2507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979247.983, "dur": 0.361,
    "args": {
      "External id": 2508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979253.570, "dur": 5.288,
    "args": {
      "External id": 2509,"Record function id": 0, "Ev Idx": 1996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979255.976, "dur": 2.048,
    "args": {
      "External id": 2510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979256.644, "dur": 1.049,
    "args": {
      "External id": 2511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979257.193, "dur": 0.365,
    "args": {
      "External id": 2512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 1999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979262.397, "dur": 4.990,
    "args": {
      "External id": 2513,"Record function id": 0, "Ev Idx": 2000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979264.294, "dur": 2.359,
    "args": {
      "External id": 2514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979264.999, "dur": 1.130,
    "args": {
      "External id": 2515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979265.550, "dur": 0.448,
    "args": {
      "External id": 2516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979270.934, "dur": 5.318,
    "args": {
      "External id": 2517,"Record function id": 0, "Ev Idx": 2004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979272.718, "dur": 2.809,
    "args": {
      "External id": 2518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979273.556, "dur": 1.446,
    "args": {
      "External id": 2519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979273.947, "dur": 0.923,
    "args": {
      "External id": 2520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979279.786, "dur": 5.595,
    "args": {
      "External id": 2521,"Record function id": 0, "Ev Idx": 2008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979281.789, "dur": 2.740,
    "args": {
      "External id": 2522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979282.547, "dur": 1.452,
    "args": {
      "External id": 2523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979283.090, "dur": 0.733,
    "args": {
      "External id": 2524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979288.874, "dur": 5.024,
    "args": {
      "External id": 2525,"Record function id": 0, "Ev Idx": 2012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979290.810, "dur": 2.210,
    "args": {
      "External id": 2526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979291.559, "dur": 0.831,
    "args": {
      "External id": 2527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979291.934, "dur": 0.329,
    "args": {
      "External id": 2528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979297.645, "dur": 5.599,
    "args": {
      "External id": 2529,"Record function id": 0, "Ev Idx": 2016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979299.750, "dur": 2.580,
    "args": {
      "External id": 2530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979300.622, "dur": 1.128,
    "args": {
      "External id": 2531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979300.996, "dur": 0.604,
    "args": {
      "External id": 2532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979306.695, "dur": 47.204,
    "args": {
      "External id": 2533,"Record function id": 0, "Ev Idx": 2020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979308.135, "dur": 42.411,
    "args": {
      "External id": 2534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979308.889, "dur": 40.918,
    "args": {
      "External id": 2535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979309.281, "dur": 39.347,
    "args": {
      "External id": 2536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979361.059, "dur": 7.277,
    "args": {
      "External id": 2537,"Record function id": 0, "Ev Idx": 2024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979365.101, "dur": 2.505,
    "args": {
      "External id": 2538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979366.129, "dur": 1.009,
    "args": {
      "External id": 2539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979366.548, "dur": 0.460,
    "args": {
      "External id": 2540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979371.724, "dur": 4.808,
    "args": {
      "External id": 2541,"Record function id": 0, "Ev Idx": 2028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979373.207, "dur": 2.564,
    "args": {
      "External id": 2542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979374.026, "dur": 1.197,
    "args": {
      "External id": 2543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979374.626, "dur": 0.464,
    "args": {
      "External id": 2544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979380.082, "dur": 4.584,
    "args": {
      "External id": 2545,"Record function id": 0, "Ev Idx": 2032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979381.630, "dur": 2.229,
    "args": {
      "External id": 2546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979382.336, "dur": 1.027,
    "args": {
      "External id": 2547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979382.768, "dur": 0.444,
    "args": {
      "External id": 2548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979388.234, "dur": 4.608,
    "args": {
      "External id": 2549,"Record function id": 0, "Ev Idx": 2036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979389.866, "dur": 2.262,
    "args": {
      "External id": 2550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979390.570, "dur": 1.012,
    "args": {
      "External id": 2551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979390.949, "dur": 0.423,
    "args": {
      "External id": 2552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979396.238, "dur": 4.565,
    "args": {
      "External id": 2553,"Record function id": 0, "Ev Idx": 2040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979397.721, "dur": 2.275,
    "args": {
      "External id": 2554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979398.449, "dur": 1.072,
    "args": {
      "External id": 2555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979398.876, "dur": 0.508,
    "args": {
      "External id": 2556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979404.128, "dur": 4.899,
    "args": {
      "External id": 2557,"Record function id": 0, "Ev Idx": 2044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979405.845, "dur": 2.362,
    "args": {
      "External id": 2558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979406.610, "dur": 1.130,
    "args": {
      "External id": 2559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979407.002, "dur": 0.616,
    "args": {
      "External id": 2560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979599.411, "dur": 16.407,
    "args": {
      "External id": 2561,"Record function id": 0, "Ev Idx": 2048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979612.090, "dur": 2.863,
    "args": {
      "External id": 2562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979613.040, "dur": 1.231,
    "args": {
      "External id": 2563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979613.469, "dur": 0.652,
    "args": {
      "External id": 2564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979619.943, "dur": 4.667,
    "args": {
      "External id": 2565,"Record function id": 0, "Ev Idx": 2052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979621.785, "dur": 2.067,
    "args": {
      "External id": 2566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979622.420, "dur": 0.917,
    "args": {
      "External id": 2567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979622.784, "dur": 0.409,
    "args": {
      "External id": 2568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979627.932, "dur": 4.871,
    "args": {
      "External id": 2569,"Record function id": 0, "Ev Idx": 2056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979629.758, "dur": 2.243,
    "args": {
      "External id": 2570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979630.598, "dur": 0.791,
    "args": {
      "External id": 2571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979630.960, "dur": 0.302,
    "args": {
      "External id": 2572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979636.004, "dur": 5.472,
    "args": {
      "External id": 2573,"Record function id": 0, "Ev Idx": 2060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979638.396, "dur": 2.343,
    "args": {
      "External id": 2574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979639.213, "dur": 1.005,
    "args": {
      "External id": 2575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979639.796, "dur": 0.295,
    "args": {
      "External id": 2576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979644.899, "dur": 4.320,
    "args": {
      "External id": 2577,"Record function id": 0, "Ev Idx": 2064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979646.585, "dur": 1.871,
    "args": {
      "External id": 2578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979647.220, "dur": 0.777,
    "args": {
      "External id": 2579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979647.567, "dur": 0.301,
    "args": {
      "External id": 2580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979652.799, "dur": 4.566,
    "args": {
      "External id": 2581,"Record function id": 0, "Ev Idx": 2068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979654.625, "dur": 1.952,
    "args": {
      "External id": 2582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979655.216, "dur": 0.858,
    "args": {
      "External id": 2583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979655.619, "dur": 0.330,
    "args": {
      "External id": 2584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979660.743, "dur": 4.409,
    "args": {
      "External id": 2585,"Record function id": 0, "Ev Idx": 2072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979662.306, "dur": 2.148,
    "args": {
      "External id": 2586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979662.962, "dur": 1.008,
    "args": {
      "External id": 2587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979663.488, "dur": 0.317,
    "args": {
      "External id": 2588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979668.451, "dur": 4.617,
    "args": {
      "External id": 2589,"Record function id": 0, "Ev Idx": 2076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979669.893, "dur": 2.381,
    "args": {
      "External id": 2590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979670.804, "dur": 0.982,
    "args": {
      "External id": 2591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979671.196, "dur": 0.430,
    "args": {
      "External id": 2592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979676.695, "dur": 5.084,
    "args": {
      "External id": 2593,"Record function id": 0, "Ev Idx": 2080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979678.156, "dur": 2.806,
    "args": {
      "External id": 2594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979679.105, "dur": 1.410,
    "args": {
      "External id": 2595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979679.506, "dur": 0.889,
    "args": {
      "External id": 2596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979685.176, "dur": 4.331,
    "args": {
      "External id": 2597,"Record function id": 0, "Ev Idx": 2084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979686.764, "dur": 2.012,
    "args": {
      "External id": 2598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979687.390, "dur": 0.919,
    "args": {
      "External id": 2599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979687.740, "dur": 0.443,
    "args": {
      "External id": 2600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979694.751, "dur": 4.493,
    "args": {
      "External id": 2601,"Record function id": 0, "Ev Idx": 2088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979696.238, "dur": 2.288,
    "args": {
      "External id": 2602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979696.847, "dur": 1.219,
    "args": {
      "External id": 2603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979697.212, "dur": 0.717,
    "args": {
      "External id": 2604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979702.720, "dur": 4.765,
    "args": {
      "External id": 2605,"Record function id": 0, "Ev Idx": 2092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979704.380, "dur": 2.314,
    "args": {
      "External id": 2606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979705.074, "dur": 1.143,
    "args": {
      "External id": 2607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979705.444, "dur": 0.640,
    "args": {
      "External id": 2608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979711.047, "dur": 43.900,
    "args": {
      "External id": 2609,"Record function id": 0, "Ev Idx": 2096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979712.920, "dur": 40.609,
    "args": {
      "External id": 2610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979713.702, "dur": 39.066,
    "args": {
      "External id": 2611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979714.082, "dur": 37.425,
    "args": {
      "External id": 2612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979765.705, "dur": 8.341,
    "args": {
      "External id": 2613,"Record function id": 0, "Ev Idx": 2100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979770.812, "dur": 2.535,
    "args": {
      "External id": 2614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979771.760, "dur": 1.100,
    "args": {
      "External id": 2615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979772.169, "dur": 0.549,
    "args": {
      "External id": 2616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979777.709, "dur": 4.709,
    "args": {
      "External id": 2617,"Record function id": 0, "Ev Idx": 2104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979779.408, "dur": 2.264,
    "args": {
      "External id": 2618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979780.223, "dur": 0.832,
    "args": {
      "External id": 2619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979780.601, "dur": 0.302,
    "args": {
      "External id": 2620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979785.988, "dur": 4.715,
    "args": {
      "External id": 2621,"Record function id": 0, "Ev Idx": 2108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979787.602, "dur": 2.333,
    "args": {
      "External id": 2622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979788.510, "dur": 0.957,
    "args": {
      "External id": 2623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979788.912, "dur": 0.422,
    "args": {
      "External id": 2624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979794.160, "dur": 5.361,
    "args": {
      "External id": 2625,"Record function id": 0, "Ev Idx": 2112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979795.993, "dur": 2.788,
    "args": {
      "External id": 2626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979796.607, "dur": 1.632,
    "args": {
      "External id": 2627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979797.666, "dur": 0.440,
    "args": {
      "External id": 2628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979802.779, "dur": 4.602,
    "args": {
      "External id": 2629,"Record function id": 0, "Ev Idx": 2116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979804.338, "dur": 2.289,
    "args": {
      "External id": 2630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979805.191, "dur": 0.960,
    "args": {
      "External id": 2631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979805.563, "dur": 0.448,
    "args": {
      "External id": 2632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979810.691, "dur": 4.370,
    "args": {
      "External id": 2633,"Record function id": 0, "Ev Idx": 2120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979812.018, "dur": 2.267,
    "args": {
      "External id": 2634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979812.939, "dur": 0.870,
    "args": {
      "External id": 2635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979813.282, "dur": 0.401,
    "args": {
      "External id": 2636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979818.300, "dur": 4.068,
    "args": {
      "External id": 2637,"Record function id": 0, "Ev Idx": 2124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979819.612, "dur": 1.989,
    "args": {
      "External id": 2638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979820.184, "dur": 0.908,
    "args": {
      "External id": 2639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979820.524, "dur": 0.444,
    "args": {
      "External id": 2640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979825.566, "dur": 4.344,
    "args": {
      "External id": 2641,"Record function id": 0, "Ev Idx": 2128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979826.994, "dur": 2.215,
    "args": {
      "External id": 2642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979827.598, "dur": 1.113,
    "args": {
      "External id": 2643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979828.157, "dur": 0.434,
    "args": {
      "External id": 2644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979833.146, "dur": 4.916,
    "args": {
      "External id": 2645,"Record function id": 0, "Ev Idx": 2132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979834.617, "dur": 2.748,
    "args": {
      "External id": 2646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979835.600, "dur": 1.204,
    "args": {
      "External id": 2647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979836.169, "dur": 0.507,
    "args": {
      "External id": 2648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979841.341, "dur": 4.519,
    "args": {
      "External id": 2649,"Record function id": 0, "Ev Idx": 2136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979842.829, "dur": 2.255,
    "args": {
      "External id": 2650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979843.780, "dur": 0.812,
    "args": {
      "External id": 2651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979844.140, "dur": 0.330,
    "args": {
      "External id": 2652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979849.218, "dur": 6.265,
    "args": {
      "External id": 2653,"Record function id": 0, "Ev Idx": 2140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979850.884, "dur": 3.673,
    "args": {
      "External id": 2654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979853.068, "dur": 0.813,
    "args": {
      "External id": 2655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979853.449, "dur": 0.302,
    "args": {
      "External id": 2656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979858.786, "dur": 5.546,
    "args": {
      "External id": 2657,"Record function id": 0, "Ev Idx": 2144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979860.850, "dur": 2.740,
    "args": {
      "External id": 2658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979861.928, "dur": 1.170,
    "args": {
      "External id": 2659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979862.523, "dur": 0.450,
    "args": {
      "External id": 2660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979867.573, "dur": 4.397,
    "args": {
      "External id": 2661,"Record function id": 0, "Ev Idx": 2148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979869.292, "dur": 1.990,
    "args": {
      "External id": 2662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979869.933, "dur": 0.821,
    "args": {
      "External id": 2663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979870.314, "dur": 0.307,
    "args": {
      "External id": 2664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979875.499, "dur": 5.710,
    "args": {
      "External id": 2665,"Record function id": 0, "Ev Idx": 2152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979877.937, "dur": 2.442,
    "args": {
      "External id": 2666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979878.642, "dur": 1.214,
    "args": {
      "External id": 2667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979879.203, "dur": 0.522,
    "args": {
      "External id": 2668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979884.884, "dur": 4.988,
    "args": {
      "External id": 2669,"Record function id": 0, "Ev Idx": 2156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979886.829, "dur": 2.216,
    "args": {
      "External id": 2670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979887.513, "dur": 1.003,
    "args": {
      "External id": 2671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979888.037, "dur": 0.348,
    "args": {
      "External id": 2672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979893.503, "dur": 4.579,
    "args": {
      "External id": 2673,"Record function id": 0, "Ev Idx": 2160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979895.183, "dur": 2.153,
    "args": {
      "External id": 2674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979895.864, "dur": 0.907,
    "args": {
      "External id": 2675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979896.220, "dur": 0.406,
    "args": {
      "External id": 2676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979901.424, "dur": 6.599,
    "args": {
      "External id": 2677,"Record function id": 0, "Ev Idx": 2164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979902.971, "dur": 4.342,
    "args": {
      "External id": 2678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979905.791, "dur": 1.043,
    "args": {
      "External id": 2679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979906.165, "dur": 0.537,
    "args": {
      "External id": 2680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979911.129, "dur": 4.771,
    "args": {
      "External id": 2681,"Record function id": 0, "Ev Idx": 2168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979912.658, "dur": 2.524,
    "args": {
      "External id": 2682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979913.406, "dur": 1.310,
    "args": {
      "External id": 2683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979913.746, "dur": 0.839,
    "args": {
      "External id": 2684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979918.985, "dur": 4.648,
    "args": {
      "External id": 2685,"Record function id": 0, "Ev Idx": 2172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979920.387, "dur": 2.412,
    "args": {
      "External id": 2686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979921.104, "dur": 1.237,
    "args": {
      "External id": 2687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979921.469, "dur": 0.744,
    "args": {
      "External id": 2688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979926.793, "dur": 4.589,
    "args": {
      "External id": 2689,"Record function id": 0, "Ev Idx": 2176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979928.425, "dur": 2.143,
    "args": {
      "External id": 2690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979929.129, "dur": 0.970,
    "args": {
      "External id": 2691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979929.486, "dur": 0.487,
    "args": {
      "External id": 2692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979934.663, "dur": 4.774,
    "args": {
      "External id": 2693,"Record function id": 0, "Ev Idx": 2180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979936.071, "dur": 2.587,
    "args": {
      "External id": 2694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979937.053, "dur": 1.126,
    "args": {
      "External id": 2695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979937.436, "dur": 0.618,
    "args": {
      "External id": 2696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979942.669, "dur": 46.295,
    "args": {
      "External id": 2697,"Record function id": 0, "Ev Idx": 2184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979944.591, "dur": 42.987,
    "args": {
      "External id": 2698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979945.456, "dur": 41.231,
    "args": {
      "External id": 2699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507979945.998, "dur": 39.509,
    "args": {
      "External id": 2700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507979999.547, "dur": 8.889,
    "args": {
      "External id": 2701,"Record function id": 0, "Ev Idx": 2188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980004.973, "dur": 2.687,
    "args": {
      "External id": 2702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980005.969, "dur": 1.229,
    "args": {
      "External id": 2703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980006.532, "dur": 0.540,
    "args": {
      "External id": 2704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980011.703, "dur": 5.070,
    "args": {
      "External id": 2705,"Record function id": 0, "Ev Idx": 2192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980013.278, "dur": 2.746,
    "args": {
      "External id": 2706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980014.236, "dur": 1.261,
    "args": {
      "External id": 2707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980014.914, "dur": 0.450,
    "args": {
      "External id": 2708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980019.948, "dur": 5.356,
    "args": {
      "External id": 2709,"Record function id": 0, "Ev Idx": 2196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980021.657, "dur": 2.861,
    "args": {
      "External id": 2710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980022.849, "dur": 1.187,
    "args": {
      "External id": 2711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980023.441, "dur": 0.470,
    "args": {
      "External id": 2712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980028.298, "dur": 47.268,
    "args": {
      "External id": 2713,"Record function id": 0, "Ev Idx": 2200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980029.783, "dur": 44.467,
    "args": {
      "External id": 2714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980030.823, "dur": 42.656,
    "args": {
      "External id": 2715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980070.601, "dur": 2.637,
    "args": {
      "External id": 2716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980092.278, "dur": 8.196,
    "args": {
      "External id": 2717,"Record function id": 0, "Ev Idx": 2204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980096.658, "dur": 3.082,
    "args": {
      "External id": 2718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980098.097, "dur": 1.074,
    "args": {
      "External id": 2719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980098.471, "dur": 0.566,
    "args": {
      "External id": 2720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980104.579, "dur": 4.802,
    "args": {
      "External id": 2721,"Record function id": 0, "Ev Idx": 2208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980106.117, "dur": 2.554,
    "args": {
      "External id": 2722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980107.213, "dur": 0.935,
    "args": {
      "External id": 2723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980107.596, "dur": 0.423,
    "args": {
      "External id": 2724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980112.761, "dur": 4.889,
    "args": {
      "External id": 2725,"Record function id": 0, "Ev Idx": 2212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980114.467, "dur": 2.386,
    "args": {
      "External id": 2726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980115.253, "dur": 1.119,
    "args": {
      "External id": 2727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980115.779, "dur": 0.463,
    "args": {
      "External id": 2728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980121.014, "dur": 4.582,
    "args": {
      "External id": 2729,"Record function id": 0, "Ev Idx": 2216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980122.629, "dur": 2.166,
    "args": {
      "External id": 2730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980123.364, "dur": 0.959,
    "args": {
      "External id": 2731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980123.737, "dur": 0.456,
    "args": {
      "External id": 2732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980129.132, "dur": 4.451,
    "args": {
      "External id": 2733,"Record function id": 0, "Ev Idx": 2220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980130.685, "dur": 2.183,
    "args": {
      "External id": 2734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980131.411, "dur": 0.999,
    "args": {
      "External id": 2735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980131.856, "dur": 0.427,
    "args": {
      "External id": 2736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980137.147, "dur": 4.937,
    "args": {
      "External id": 2737,"Record function id": 0, "Ev Idx": 2224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980138.878, "dur": 2.394,
    "args": {
      "External id": 2738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980139.743, "dur": 0.917,
    "args": {
      "External id": 2739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980140.074, "dur": 0.453,
    "args": {
      "External id": 2740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980145.577, "dur": 4.579,
    "args": {
      "External id": 2741,"Record function id": 0, "Ev Idx": 2228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980147.175, "dur": 2.257,
    "args": {
      "External id": 2742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980148.163, "dur": 0.819,
    "args": {
      "External id": 2743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980148.536, "dur": 0.319,
    "args": {
      "External id": 2744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980153.838, "dur": 5.293,
    "args": {
      "External id": 2745,"Record function id": 0, "Ev Idx": 2232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980156.418, "dur": 1.983,
    "args": {
      "External id": 2746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980157.093, "dur": 0.893,
    "args": {
      "External id": 2747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980157.530, "dur": 0.309,
    "args": {
      "External id": 2748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980162.661, "dur": 4.441,
    "args": {
      "External id": 2749,"Record function id": 0, "Ev Idx": 2236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980164.089, "dur": 2.302,
    "args": {
      "External id": 2750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980165.043, "dur": 0.859,
    "args": {
      "External id": 2751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980165.388, "dur": 0.390,
    "args": {
      "External id": 2752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980170.890, "dur": 6.669,
    "args": {
      "External id": 2753,"Record function id": 0, "Ev Idx": 2240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980172.670, "dur": 4.073,
    "args": {
      "External id": 2754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980173.717, "dur": 2.549,
    "args": {
      "External id": 2755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980174.140, "dur": 1.935,
    "args": {
      "External id": 2756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980180.989, "dur": 4.339,
    "args": {
      "External id": 2757,"Record function id": 0, "Ev Idx": 2244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980182.626, "dur": 2.011,
    "args": {
      "External id": 2758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980183.347, "dur": 0.844,
    "args": {
      "External id": 2759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980183.674, "dur": 0.372,
    "args": {
      "External id": 2760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980188.820, "dur": 4.659,
    "args": {
      "External id": 2761,"Record function id": 0, "Ev Idx": 2248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980190.515, "dur": 2.160,
    "args": {
      "External id": 2762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980191.229, "dur": 1.016,
    "args": {
      "External id": 2763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980191.663, "dur": 0.455,
    "args": {
      "External id": 2764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980196.725, "dur": 4.326,
    "args": {
      "External id": 2765,"Record function id": 0, "Ev Idx": 2252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980198.231, "dur": 2.112,
    "args": {
      "External id": 2766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980198.915, "dur": 0.958,
    "args": {
      "External id": 2767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980199.282, "dur": 0.460,
    "args": {
      "External id": 2768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980204.409, "dur": 5.008,
    "args": {
      "External id": 2769,"Record function id": 0, "Ev Idx": 2256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980206.108, "dur": 2.461,
    "args": {
      "External id": 2770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980206.809, "dur": 1.277,
    "args": {
      "External id": 2771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980207.138, "dur": 0.820,
    "args": {
      "External id": 2772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980212.674, "dur": 4.687,
    "args": {
      "External id": 2773,"Record function id": 0, "Ev Idx": 2260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980214.193, "dur": 2.368,
    "args": {
      "External id": 2774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980214.915, "dur": 1.203,
    "args": {
      "External id": 2775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980215.282, "dur": 0.714,
    "args": {
      "External id": 2776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980220.497, "dur": 45.902,
    "args": {
      "External id": 2777,"Record function id": 0, "Ev Idx": 2264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980222.124, "dur": 42.859,
    "args": {
      "External id": 2778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 2265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980222.914, "dur": 41.243,
    "args": {
      "External id": 2779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 2266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980223.280, "dur": 39.645,
    "args": {
      "External id": 2780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 2267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980276.110, "dur": 7.112,
    "args": {
      "External id": 2781,"Record function id": 0, "Ev Idx": 2268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980280.235, "dur": 2.255,
    "args": {
      "External id": 2782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980281.035, "dur": 1.010,
    "args": {
      "External id": 2783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980281.475, "dur": 0.442,
    "args": {
      "External id": 2784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980286.852, "dur": 4.559,
    "args": {
      "External id": 2785,"Record function id": 0, "Ev Idx": 2272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980288.534, "dur": 2.159,
    "args": {
      "External id": 2786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980289.264, "dur": 0.968,
    "args": {
      "External id": 2787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980289.630, "dur": 0.472,
    "args": {
      "External id": 2788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980294.542, "dur": 4.567,
    "args": {
      "External id": 2789,"Record function id": 0, "Ev Idx": 2276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980296.279, "dur": 2.104,
    "args": {
      "External id": 2790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980297.160, "dur": 0.899,
    "args": {
      "External id": 2791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980297.495, "dur": 0.441,
    "args": {
      "External id": 2792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980302.436, "dur": 5.306,
    "args": {
      "External id": 2793,"Record function id": 0, "Ev Idx": 2280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980304.323, "dur": 2.445,
    "args": {
      "External id": 2794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980305.081, "dur": 1.084,
    "args": {
      "External id": 2795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980305.584, "dur": 0.448,
    "args": {
      "External id": 2796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980311.236, "dur": 4.517,
    "args": {
      "External id": 2797,"Record function id": 0, "Ev Idx": 2284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980313.045, "dur": 1.999,
    "args": {
      "External id": 2798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980313.764, "dur": 0.822,
    "args": {
      "External id": 2799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980314.127, "dur": 0.320,
    "args": {
      "External id": 2800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980319.170, "dur": 4.288,
    "args": {
      "External id": 2801,"Record function id": 0, "Ev Idx": 2288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980320.711, "dur": 2.058,
    "args": {
      "External id": 2802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 2289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980321.346, "dur": 0.920,
    "args": {
      "External id": 2803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 2290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980321.710, "dur": 0.426,
    "args": {
      "External id": 2804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 2291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980326.735, "dur": 5.466,
    "args": {
      "External id": 2805,"Record function id": 0, "Ev Idx": 2292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980328.646, "dur": 2.792,
    "args": {
      "External id": 2806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980329.558, "dur": 1.152,
    "args": {
      "External id": 2807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980330.070, "dur": 0.448,
    "args": {
      "External id": 2808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980335.613, "dur": 5.454,
    "args": {
      "External id": 2809,"Record function id": 0, "Ev Idx": 2296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980337.517, "dur": 2.789,
    "args": {
      "External id": 2810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980338.614, "dur": 1.001,
    "args": {
      "External id": 2811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980338.985, "dur": 0.490,
    "args": {
      "External id": 2812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980344.265, "dur": 5.028,
    "args": {
      "External id": 2813,"Record function id": 0, "Ev Idx": 2300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980346.308, "dur": 2.115,
    "args": {
      "External id": 2814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980346.983, "dur": 0.940,
    "args": {
      "External id": 2815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980347.373, "dur": 0.421,
    "args": {
      "External id": 2816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980352.754, "dur": 4.635,
    "args": {
      "External id": 2817,"Record function id": 0, "Ev Idx": 2304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980354.563, "dur": 2.102,
    "args": {
      "External id": 2818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980355.328, "dur": 0.829,
    "args": {
      "External id": 2819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980355.709, "dur": 0.320,
    "args": {
      "External id": 2820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980360.373, "dur": 4.909,
    "args": {
      "External id": 2821,"Record function id": 0, "Ev Idx": 2308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980362.242, "dur": 2.234,
    "args": {
      "External id": 2822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980363.151, "dur": 0.841,
    "args": {
      "External id": 2823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980363.522, "dur": 0.326,
    "args": {
      "External id": 2824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980368.547, "dur": 4.558,
    "args": {
      "External id": 2825,"Record function id": 0, "Ev Idx": 2312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980370.137, "dur": 2.260,
    "args": {
      "External id": 2826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980370.891, "dur": 0.952,
    "args": {
      "External id": 2827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980371.271, "dur": 0.423,
    "args": {
      "External id": 2828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980375.964, "dur": 4.463,
    "args": {
      "External id": 2829,"Record function id": 0, "Ev Idx": 2316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980377.373, "dur": 2.315,
    "args": {
      "External id": 2830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980378.108, "dur": 1.062,
    "args": {
      "External id": 2831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980378.527, "dur": 0.506,
    "args": {
      "External id": 2832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507980423.254, "dur": 236.322,
    "args": {
      "External id": 2833,"Record function id": 0, "Sequence number": 4688, "Fwd thread id": 1, "Ev Idx": 2320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507980432.059, "dur": 181.149,
    "args": {
      "External id": 2834,"Sequence number": 4688, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 2321
    }
  },
  {
    "ph": "f", "id": 6, "pid": 14565, "tid": 14640, "ts": 1290507980432.059,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290507980517.346, "dur": 81.265,
    "args": {
      "External id": 2835,"Record function id": 0, "scope": "user_scope", "Ev Idx": 2322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290507980625.301, "dur": 31.348,
    "args": {
      "External id": 2836,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 2323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980666.032, "dur": 6.508,
    "args": {
      "External id": 2837,"Record function id": 0, "Ev Idx": 2324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507980668.007, "dur": 3.635,
    "args": {
      "External id": 2838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 2325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980669.269, "dur": 1.984,
    "args": {
      "External id": 2839,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 2326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507980670.210, "dur": 0.861,
    "args": {
      "External id": 2840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 2327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507983627.293, "dur": 2370.647,
    "args": {
      "External id": 2841,"Record function id": 0, "Sequence number": 4691, "Fwd thread id": 1, "Ev Idx": 2328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507983629.830, "dur": 2274.204,
    "args": {
      "External id": 2842,"Sequence number": 4691, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 2329
    }
  },
  {
    "ph": "f", "id": 7, "pid": 14565, "tid": 14640, "ts": 1290507983629.830,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290507984096.878, "dur": 1780.872,
    "args": {
      "External id": 2843,"Record function id": 0, "scope": "user_scope", "Ev Idx": 2330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290507984807.882, "dur": 31.345,
    "args": {
      "External id": 2844,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 2331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986011.455, "dur": 10.313,
    "args": {
      "External id": 2845,"Record function id": 0, "Ev Idx": 2332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986016.691, "dur": 3.876,
    "args": {
      "External id": 2846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986018.130, "dur": 1.959,
    "args": {
      "External id": 2847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986018.997, "dur": 0.927,
    "args": {
      "External id": 2848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986025.536, "dur": 4.994,
    "args": {
      "External id": 2849,"Record function id": 0, "Ev Idx": 2336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986027.145, "dur": 2.515,
    "args": {
      "External id": 2850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986028.149, "dur": 1.193,
    "args": {
      "External id": 2851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986028.700, "dur": 0.478,
    "args": {
      "External id": 2852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986033.981, "dur": 4.403,
    "args": {
      "External id": 2853,"Record function id": 0, "Ev Idx": 2340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986035.383, "dur": 2.203,
    "args": {
      "External id": 2854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986036.103, "dur": 1.171,
    "args": {
      "External id": 2855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986036.648, "dur": 0.501,
    "args": {
      "External id": 2856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986041.733, "dur": 4.268,
    "args": {
      "External id": 2857,"Record function id": 0, "Ev Idx": 2344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986043.163, "dur": 2.031,
    "args": {
      "External id": 2858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986043.888, "dur": 0.935,
    "args": {
      "External id": 2859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986044.276, "dur": 0.375,
    "args": {
      "External id": 2860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986049.280, "dur": 4.107,
    "args": {
      "External id": 2861,"Record function id": 0, "Ev Idx": 2348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986050.604, "dur": 2.034,
    "args": {
      "External id": 2862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986051.328, "dur": 0.981,
    "args": {
      "External id": 2863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986051.712, "dur": 0.432,
    "args": {
      "External id": 2864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986056.786, "dur": 3.973,
    "args": {
      "External id": 2865,"Record function id": 0, "Ev Idx": 2352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986058.063, "dur": 1.957,
    "args": {
      "External id": 2866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986058.767, "dur": 0.949,
    "args": {
      "External id": 2867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986059.175, "dur": 0.411,
    "args": {
      "External id": 2868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986064.025, "dur": 4.458,
    "args": {
      "External id": 2869,"Record function id": 0, "Ev Idx": 2356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986065.420, "dur": 2.338,
    "args": {
      "External id": 2870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986066.166, "dur": 1.282,
    "args": {
      "External id": 2871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986066.767, "dur": 0.532,
    "args": {
      "External id": 2872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986071.921, "dur": 4.208,
    "args": {
      "External id": 2873,"Record function id": 0, "Ev Idx": 2360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986073.412, "dur": 1.942,
    "args": {
      "External id": 2874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986074.051, "dur": 0.994,
    "args": {
      "External id": 2875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986074.541, "dur": 0.375,
    "args": {
      "External id": 2876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986079.539, "dur": 10.648,
    "args": {
      "External id": 2877,"Record function id": 0, "Ev Idx": 2364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986087.361, "dur": 2.074,
    "args": {
      "External id": 2878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986088.028, "dur": 1.089,
    "args": {
      "External id": 2879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986088.458, "dur": 0.529,
    "args": {
      "External id": 2880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986093.562, "dur": 5.442,
    "args": {
      "External id": 2881,"Record function id": 0, "Ev Idx": 2368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986095.152, "dur": 3.002,
    "args": {
      "External id": 2882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986096.137, "dur": 1.014,
    "args": {
      "External id": 2883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986096.635, "dur": 0.365,
    "args": {
      "External id": 2884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986102.429, "dur": 4.345,
    "args": {
      "External id": 2885,"Record function id": 0, "Ev Idx": 2372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986103.723, "dur": 2.244,
    "args": {
      "External id": 2886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986104.503, "dur": 1.043,
    "args": {
      "External id": 2887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986105.074, "dur": 0.332,
    "args": {
      "External id": 2888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986110.020, "dur": 4.272,
    "args": {
      "External id": 2889,"Record function id": 0, "Ev Idx": 2376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986111.302, "dur": 2.172,
    "args": {
      "External id": 2890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986111.964, "dur": 1.188,
    "args": {
      "External id": 2891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986112.638, "dur": 0.381,
    "args": {
      "External id": 2892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986117.556, "dur": 4.401,
    "args": {
      "External id": 2893,"Record function id": 0, "Ev Idx": 2380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986118.997, "dur": 2.158,
    "args": {
      "External id": 2894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986119.765, "dur": 1.109,
    "args": {
      "External id": 2895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986120.374, "dur": 0.361,
    "args": {
      "External id": 2896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986125.388, "dur": 5.174,
    "args": {
      "External id": 2897,"Record function id": 0, "Ev Idx": 2384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986127.586, "dur": 2.183,
    "args": {
      "External id": 2898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986128.316, "dur": 1.161,
    "args": {
      "External id": 2899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986128.959, "dur": 0.378,
    "args": {
      "External id": 2900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986133.926, "dur": 3.975,
    "args": {
      "External id": 2901,"Record function id": 0, "Ev Idx": 2388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986135.331, "dur": 1.816,
    "args": {
      "External id": 2902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986135.908, "dur": 0.931,
    "args": {
      "External id": 2903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986136.303, "dur": 0.390,
    "args": {
      "External id": 2904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986141.614, "dur": 4.149,
    "args": {
      "External id": 2905,"Record function id": 0, "Ev Idx": 2392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986143.193, "dur": 1.822,
    "args": {
      "External id": 2906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986143.817, "dur": 0.887,
    "args": {
      "External id": 2907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986144.224, "dur": 0.335,
    "args": {
      "External id": 2908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986149.440, "dur": 104.663,
    "args": {
      "External id": 2909,"Record function id": 0, "Ev Idx": 2396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986150.969, "dur": 97.912,
    "args": {
      "External id": 2910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986151.687, "dur": 92.866,
    "args": {
      "External id": 2911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986152.095, "dur": 89.361,
    "args": {
      "External id": 2912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986282.003, "dur": 13.648,
    "args": {
      "External id": 2913,"Record function id": 0, "Ev Idx": 2400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986289.464, "dur": 5.455,
    "args": {
      "External id": 2914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986291.731, "dur": 2.664,
    "args": {
      "External id": 2915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986293.169, "dur": 1.065,
    "args": {
      "External id": 2916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986300.174, "dur": 5.894,
    "args": {
      "External id": 2917,"Record function id": 0, "Ev Idx": 2404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986302.653, "dur": 2.701,
    "args": {
      "External id": 2918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986303.601, "dur": 1.244,
    "args": {
      "External id": 2919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986304.358, "dur": 0.341,
    "args": {
      "External id": 2920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986309.498, "dur": 6.189,
    "args": {
      "External id": 2921,"Record function id": 0, "Ev Idx": 2408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986311.565, "dur": 3.268,
    "args": {
      "External id": 2922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986312.625, "dur": 1.746,
    "args": {
      "External id": 2923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986313.652, "dur": 0.585,
    "args": {
      "External id": 2924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986319.329, "dur": 6.464,
    "args": {
      "External id": 2925,"Record function id": 0, "Ev Idx": 2412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986321.782, "dur": 3.223,
    "args": {
      "External id": 2926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986323.059, "dur": 0.970,
    "args": {
      "External id": 2927,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986323.543, "dur": 0.329,
    "args": {
      "External id": 2928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986329.530, "dur": 5.039,
    "args": {
      "External id": 2929,"Record function id": 0, "Ev Idx": 2416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986331.550, "dur": 2.280,
    "args": {
      "External id": 2930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986332.410, "dur": 0.772,
    "args": {
      "External id": 2931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986332.777, "dur": 0.281,
    "args": {
      "External id": 2932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986338.130, "dur": 5.180,
    "args": {
      "External id": 2933,"Record function id": 0, "Ev Idx": 2420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986340.245, "dur": 2.232,
    "args": {
      "External id": 2934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986340.927, "dur": 1.075,
    "args": {
      "External id": 2935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986341.346, "dur": 0.531,
    "args": {
      "External id": 2936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986346.760, "dur": 5.080,
    "args": {
      "External id": 2937,"Record function id": 0, "Ev Idx": 2424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986348.804, "dur": 2.310,
    "args": {
      "External id": 2938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986349.476, "dur": 1.165,
    "args": {
      "External id": 2939,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986349.840, "dur": 0.333,
    "args": {
      "External id": 2940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986355.210, "dur": 4.593,
    "args": {
      "External id": 2941,"Record function id": 0, "Ev Idx": 2428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986356.740, "dur": 2.239,
    "args": {
      "External id": 2942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986357.439, "dur": 0.891,
    "args": {
      "External id": 2943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986357.851, "dur": 0.342,
    "args": {
      "External id": 2944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986363.232, "dur": 4.672,
    "args": {
      "External id": 2945,"Record function id": 0, "Ev Idx": 2432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986365.126, "dur": 2.066,
    "args": {
      "External id": 2946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986365.905, "dur": 0.811,
    "args": {
      "External id": 2947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986366.272, "dur": 0.310,
    "args": {
      "External id": 2948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986371.309, "dur": 4.415,
    "args": {
      "External id": 2949,"Record function id": 0, "Ev Idx": 2436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986372.809, "dur": 2.089,
    "args": {
      "External id": 2950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986373.534, "dur": 0.821,
    "args": {
      "External id": 2951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986373.905, "dur": 0.324,
    "args": {
      "External id": 2952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986379.225, "dur": 5.063,
    "args": {
      "External id": 2953,"Record function id": 0, "Ev Idx": 2440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986381.193, "dur": 2.376,
    "args": {
      "External id": 2954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986382.188, "dur": 0.889,
    "args": {
      "External id": 2955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986382.560, "dur": 0.388,
    "args": {
      "External id": 2956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986387.724, "dur": 4.731,
    "args": {
      "External id": 2957,"Record function id": 0, "Ev Idx": 2444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986389.780, "dur": 1.969,
    "args": {
      "External id": 2958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986390.460, "dur": 0.802,
    "args": {
      "External id": 2959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986390.814, "dur": 0.318,
    "args": {
      "External id": 2960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986395.926, "dur": 6.051,
    "args": {
      "External id": 2961,"Record function id": 0, "Ev Idx": 2448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986398.517, "dur": 2.620,
    "args": {
      "External id": 2962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986399.209, "dur": 1.431,
    "args": {
      "External id": 2963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986399.590, "dur": 0.920,
    "args": {
      "External id": 2964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986405.692, "dur": 4.601,
    "args": {
      "External id": 2965,"Record function id": 0, "Ev Idx": 2452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986407.459, "dur": 2.002,
    "args": {
      "External id": 2966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986408.059, "dur": 0.849,
    "args": {
      "External id": 2967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986408.427, "dur": 0.349,
    "args": {
      "External id": 2968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986413.800, "dur": 5.074,
    "args": {
      "External id": 2969,"Record function id": 0, "Ev Idx": 2456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986415.633, "dur": 2.472,
    "args": {
      "External id": 2970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986416.692, "dur": 0.809,
    "args": {
      "External id": 2971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986417.053, "dur": 0.314,
    "args": {
      "External id": 2972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986422.355, "dur": 4.772,
    "args": {
      "External id": 2973,"Record function id": 0, "Ev Idx": 2460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986424.190, "dur": 2.204,
    "args": {
      "External id": 2974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986424.992, "dur": 0.814,
    "args": {
      "External id": 2975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986425.359, "dur": 0.319,
    "args": {
      "External id": 2976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986430.807, "dur": 5.079,
    "args": {
      "External id": 2977,"Record function id": 0, "Ev Idx": 2464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986432.296, "dur": 2.755,
    "args": {
      "External id": 2978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986433.024, "dur": 1.480,
    "args": {
      "External id": 2979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986433.540, "dur": 0.819,
    "args": {
      "External id": 2980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986443.081, "dur": 5.230,
    "args": {
      "External id": 2981,"Record function id": 0, "Ev Idx": 2468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986444.555, "dur": 3.041,
    "args": {
      "External id": 2982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986445.503, "dur": 1.542,
    "args": {
      "External id": 2983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986445.886, "dur": 1.022,
    "args": {
      "External id": 2984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986451.900, "dur": 117.980,
    "args": {
      "External id": 2985,"Record function id": 0, "Ev Idx": 2472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986453.755, "dur": 110.122,
    "args": {
      "External id": 2986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986454.457, "dur": 106.586,
    "args": {
      "External id": 2987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986454.828, "dur": 99.765,
    "args": {
      "External id": 2988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986602.018, "dur": 15.382,
    "args": {
      "External id": 2989,"Record function id": 0, "Ev Idx": 2476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986610.171, "dur": 6.275,
    "args": {
      "External id": 2990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986613.538, "dur": 2.359,
    "args": {
      "External id": 2991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986614.610, "dur": 1.106,
    "args": {
      "External id": 2992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986621.701, "dur": 5.689,
    "args": {
      "External id": 2993,"Record function id": 0, "Ev Idx": 2480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986623.713, "dur": 2.800,
    "args": {
      "External id": 2994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986624.584, "dur": 1.400,
    "args": {
      "External id": 2995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986625.398, "dur": 0.426,
    "args": {
      "External id": 2996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986631.150, "dur": 5.000,
    "args": {
      "External id": 2997,"Record function id": 0, "Ev Idx": 2484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986632.915, "dur": 2.490,
    "args": {
      "External id": 2998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986633.740, "dur": 1.150,
    "args": {
      "External id": 2999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986634.330, "dur": 0.431,
    "args": {
      "External id": 3000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986639.684, "dur": 5.213,
    "args": {
      "External id": 3001,"Record function id": 0, "Ev Idx": 2488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986641.730, "dur": 2.308,
    "args": {
      "External id": 3002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986642.562, "dur": 0.925,
    "args": {
      "External id": 3003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986642.955, "dur": 0.399,
    "args": {
      "External id": 3004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986648.304, "dur": 5.256,
    "args": {
      "External id": 3005,"Record function id": 0, "Ev Idx": 2492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986650.006, "dur": 2.698,
    "args": {
      "External id": 3006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986651.114, "dur": 1.062,
    "args": {
      "External id": 3007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986651.661, "dur": 0.383,
    "args": {
      "External id": 3008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986657.096, "dur": 4.817,
    "args": {
      "External id": 3009,"Record function id": 0, "Ev Idx": 2496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986658.838, "dur": 2.226,
    "args": {
      "External id": 3010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986659.669, "dur": 0.866,
    "args": {
      "External id": 3011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986660.039, "dur": 0.344,
    "args": {
      "External id": 3012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986665.400, "dur": 5.832,
    "args": {
      "External id": 3013,"Record function id": 0, "Ev Idx": 2500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986667.412, "dur": 2.932,
    "args": {
      "External id": 3014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986668.377, "dur": 0.866,
    "args": {
      "External id": 3015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986668.746, "dur": 0.354,
    "args": {
      "External id": 3016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986674.657, "dur": 4.737,
    "args": {
      "External id": 3017,"Record function id": 0, "Ev Idx": 2504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986676.029, "dur": 2.478,
    "args": {
      "External id": 3018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986676.953, "dur": 0.978,
    "args": {
      "External id": 3019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986677.458, "dur": 0.338,
    "args": {
      "External id": 3020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986682.878, "dur": 4.741,
    "args": {
      "External id": 3021,"Record function id": 0, "Ev Idx": 2508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986684.804, "dur": 2.067,
    "args": {
      "External id": 3022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986685.529, "dur": 0.847,
    "args": {
      "External id": 3023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986685.899, "dur": 0.342,
    "args": {
      "External id": 3024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986690.919, "dur": 4.617,
    "args": {
      "External id": 3025,"Record function id": 0, "Ev Idx": 2512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986692.337, "dur": 2.456,
    "args": {
      "External id": 3026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986693.222, "dur": 1.055,
    "args": {
      "External id": 3027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986693.651, "dur": 0.491,
    "args": {
      "External id": 3028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986698.814, "dur": 4.457,
    "args": {
      "External id": 3029,"Record function id": 0, "Ev Idx": 2516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986700.282, "dur": 2.260,
    "args": {
      "External id": 3030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986700.987, "dur": 0.897,
    "args": {
      "External id": 3031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986701.372, "dur": 0.380,
    "args": {
      "External id": 3032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986706.494, "dur": 4.667,
    "args": {
      "External id": 3033,"Record function id": 0, "Ev Idx": 2520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986708.154, "dur": 2.148,
    "args": {
      "External id": 3034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986708.861, "dur": 0.916,
    "args": {
      "External id": 3035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986709.242, "dur": 0.404,
    "args": {
      "External id": 3036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986714.408, "dur": 4.242,
    "args": {
      "External id": 3037,"Record function id": 0, "Ev Idx": 2524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986715.743, "dur": 2.165,
    "args": {
      "External id": 3038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986716.473, "dur": 0.893,
    "args": {
      "External id": 3039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986716.843, "dur": 0.393,
    "args": {
      "External id": 3040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986721.831, "dur": 6.193,
    "args": {
      "External id": 3041,"Record function id": 0, "Ev Idx": 2528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986724.630, "dur": 2.674,
    "args": {
      "External id": 3042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986725.348, "dur": 1.420,
    "args": {
      "External id": 3043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986725.723, "dur": 0.912,
    "args": {
      "External id": 3044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986731.383, "dur": 4.720,
    "args": {
      "External id": 3045,"Record function id": 0, "Ev Idx": 2532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986733.098, "dur": 2.158,
    "args": {
      "External id": 3046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986733.826, "dur": 0.903,
    "args": {
      "External id": 3047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986734.189, "dur": 0.402,
    "args": {
      "External id": 3048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986739.487, "dur": 5.016,
    "args": {
      "External id": 3049,"Record function id": 0, "Ev Idx": 2536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986741.380, "dur": 2.273,
    "args": {
      "External id": 3050,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986742.073, "dur": 1.055,
    "args": {
      "External id": 3051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986742.595, "dur": 0.392,
    "args": {
      "External id": 3052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986747.782, "dur": 4.755,
    "args": {
      "External id": 3053,"Record function id": 0, "Ev Idx": 2540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986749.431, "dur": 2.256,
    "args": {
      "External id": 3054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986750.293, "dur": 0.874,
    "args": {
      "External id": 3055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986750.655, "dur": 0.368,
    "args": {
      "External id": 3056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986755.844, "dur": 5.825,
    "args": {
      "External id": 3057,"Record function id": 0, "Ev Idx": 2544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986757.677, "dur": 3.093,
    "args": {
      "External id": 3058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986758.551, "dur": 1.720,
    "args": {
      "External id": 3059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986759.166, "dur": 0.963,
    "args": {
      "External id": 3060,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986765.151, "dur": 5.809,
    "args": {
      "External id": 3061,"Record function id": 0, "Ev Idx": 2548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986766.961, "dur": 3.102,
    "args": {
      "External id": 3062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986768.255, "dur": 1.386,
    "args": {
      "External id": 3063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986768.641, "dur": 0.862,
    "args": {
      "External id": 3064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986774.619, "dur": 4.503,
    "args": {
      "External id": 3065,"Record function id": 0, "Ev Idx": 2552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986776.227, "dur": 2.143,
    "args": {
      "External id": 3066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986776.972, "dur": 0.898,
    "args": {
      "External id": 3067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986777.349, "dur": 0.381,
    "args": {
      "External id": 3068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986782.413, "dur": 4.742,
    "args": {
      "External id": 3069,"Record function id": 0, "Ev Idx": 2556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507986784.012, "dur": 2.429,
    "args": {
      "External id": 3070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986784.800, "dur": 1.119,
    "args": {
      "External id": 3071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507986785.167, "dur": 0.604,
    "args": {
      "External id": 3072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987146.734, "dur": 29.936,
    "args": {
      "External id": 3073,"Record function id": 0, "Ev Idx": 2560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987168.404, "dur": 6.808,
    "args": {
      "External id": 3074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987170.446, "dur": 4.036,
    "args": {
      "External id": 3075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987171.616, "dur": 2.634,
    "args": {
      "External id": 3076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987187.125, "dur": 5.799,
    "args": {
      "External id": 3077,"Record function id": 0, "Ev Idx": 2564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987189.624, "dur": 2.560,
    "args": {
      "External id": 3078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987190.346, "dur": 1.453,
    "args": {
      "External id": 3079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987191.125, "dur": 0.549,
    "args": {
      "External id": 3080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987197.094, "dur": 5.227,
    "args": {
      "External id": 3081,"Record function id": 0, "Ev Idx": 2568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987199.289, "dur": 2.274,
    "args": {
      "External id": 3082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987199.970, "dur": 1.203,
    "args": {
      "External id": 3083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987200.723, "dur": 0.320,
    "args": {
      "External id": 3084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987206.169, "dur": 5.211,
    "args": {
      "External id": 3085,"Record function id": 0, "Ev Idx": 2572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987208.366, "dur": 2.317,
    "args": {
      "External id": 3086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987209.053, "dur": 1.137,
    "args": {
      "External id": 3087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987209.611, "dur": 0.452,
    "args": {
      "External id": 3088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987215.150, "dur": 4.465,
    "args": {
      "External id": 3089,"Record function id": 0, "Ev Idx": 2576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987216.957, "dur": 1.845,
    "args": {
      "External id": 3090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987217.522, "dur": 0.892,
    "args": {
      "External id": 3091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987217.978, "dur": 0.311,
    "args": {
      "External id": 3092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987223.169, "dur": 4.272,
    "args": {
      "External id": 3093,"Record function id": 0, "Ev Idx": 2580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987224.886, "dur": 1.783,
    "args": {
      "External id": 3094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987225.472, "dur": 0.824,
    "args": {
      "External id": 3095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987225.856, "dur": 0.318,
    "args": {
      "External id": 3096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987230.818, "dur": 5.999,
    "args": {
      "External id": 3097,"Record function id": 0, "Ev Idx": 2584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987232.387, "dur": 3.683,
    "args": {
      "External id": 3098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987232.989, "dur": 2.666,
    "args": {
      "External id": 3099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987235.148, "dur": 0.340,
    "args": {
      "External id": 3100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987240.497, "dur": 7.036,
    "args": {
      "External id": 3101,"Record function id": 0, "Ev Idx": 2588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987244.021, "dur": 2.671,
    "args": {
      "External id": 3102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987244.954, "dur": 0.798,
    "args": {
      "External id": 3103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987245.324, "dur": 0.298,
    "args": {
      "External id": 3104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987250.900, "dur": 4.593,
    "args": {
      "External id": 3105,"Record function id": 0, "Ev Idx": 2592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987252.436, "dur": 2.340,
    "args": {
      "External id": 3106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987253.446, "dur": 0.816,
    "args": {
      "External id": 3107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987253.824, "dur": 0.304,
    "args": {
      "External id": 3108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987258.863, "dur": 3.836,
    "args": {
      "External id": 3109,"Record function id": 0, "Ev Idx": 2596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987260.140, "dur": 1.816,
    "args": {
      "External id": 3110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987260.716, "dur": 0.847,
    "args": {
      "External id": 3111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987261.135, "dur": 0.298,
    "args": {
      "External id": 3112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987266.740, "dur": 4.106,
    "args": {
      "External id": 3113,"Record function id": 0, "Ev Idx": 2600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987268.202, "dur": 1.840,
    "args": {
      "External id": 3114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987268.874, "dur": 0.787,
    "args": {
      "External id": 3115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987269.236, "dur": 0.298,
    "args": {
      "External id": 3116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987274.331, "dur": 3.868,
    "args": {
      "External id": 3117,"Record function id": 0, "Ev Idx": 2604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987275.606, "dur": 1.833,
    "args": {
      "External id": 3118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987276.217, "dur": 0.820,
    "args": {
      "External id": 3119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987276.598, "dur": 0.310,
    "args": {
      "External id": 3120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987282.047, "dur": 3.970,
    "args": {
      "External id": 3121,"Record function id": 0, "Ev Idx": 2608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987283.497, "dur": 1.800,
    "args": {
      "External id": 3122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987284.041, "dur": 0.880,
    "args": {
      "External id": 3123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987284.431, "dur": 0.348,
    "args": {
      "External id": 3124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987289.722, "dur": 7.047,
    "args": {
      "External id": 3125,"Record function id": 0, "Ev Idx": 2612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987292.571, "dur": 3.409,
    "args": {
      "External id": 3126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987294.167, "dur": 1.380,
    "args": {
      "External id": 3127,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987294.661, "dur": 0.755,
    "args": {
      "External id": 3128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987300.813, "dur": 4.364,
    "args": {
      "External id": 3129,"Record function id": 0, "Ev Idx": 2616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987302.518, "dur": 1.883,
    "args": {
      "External id": 3130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987303.104, "dur": 0.910,
    "args": {
      "External id": 3131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987303.509, "dur": 0.367,
    "args": {
      "External id": 3132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987308.964, "dur": 4.752,
    "args": {
      "External id": 3133,"Record function id": 0, "Ev Idx": 2620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987310.648, "dur": 2.270,
    "args": {
      "External id": 3134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987311.448, "dur": 0.820,
    "args": {
      "External id": 3135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987311.816, "dur": 0.326,
    "args": {
      "External id": 3136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987317.700, "dur": 4.810,
    "args": {
      "External id": 3137,"Record function id": 0, "Ev Idx": 2624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987319.313, "dur": 2.398,
    "args": {
      "External id": 3138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987319.978, "dur": 0.927,
    "args": {
      "External id": 3139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987320.354, "dur": 0.412,
    "args": {
      "External id": 3140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987326.070, "dur": 84.146,
    "args": {
      "External id": 3141,"Record function id": 0, "Ev Idx": 2628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987327.349, "dur": 79.699,
    "args": {
      "External id": 3142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987327.896, "dur": 76.727,
    "args": {
      "External id": 3143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987328.284, "dur": 74.268,
    "args": {
      "External id": 3144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987429.192, "dur": 14.412,
    "args": {
      "External id": 3145,"Record function id": 0, "Ev Idx": 2632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987437.151, "dur": 5.633,
    "args": {
      "External id": 3146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987439.382, "dur": 2.378,
    "args": {
      "External id": 3147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987440.751, "dur": 0.842,
    "args": {
      "External id": 3148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987447.683, "dur": 5.338,
    "args": {
      "External id": 3149,"Record function id": 0, "Ev Idx": 2636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987449.691, "dur": 2.535,
    "args": {
      "External id": 3150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987450.561, "dur": 1.238,
    "args": {
      "External id": 3151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987451.300, "dur": 0.373,
    "args": {
      "External id": 3152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987456.571, "dur": 93.395,
    "args": {
      "External id": 3153,"Record function id": 0, "Ev Idx": 2640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987459.299, "dur": 87.895,
    "args": {
      "External id": 3154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987533.263, "dur": 12.337,
    "args": {
      "External id": 3155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987538.255, "dur": 7.109,
    "args": {
      "External id": 3156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987561.257, "dur": 6.581,
    "args": {
      "External id": 3157,"Record function id": 0, "Ev Idx": 2644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987563.468, "dur": 3.050,
    "args": {
      "External id": 3158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987564.492, "dur": 1.386,
    "args": {
      "External id": 3159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987565.154, "dur": 0.591,
    "args": {
      "External id": 3160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987573.297, "dur": 7.734,
    "args": {
      "External id": 3161,"Record function id": 0, "Ev Idx": 2648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987576.142, "dur": 3.597,
    "args": {
      "External id": 3162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987577.414, "dur": 1.698,
    "args": {
      "External id": 3163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987578.156, "dur": 0.675,
    "args": {
      "External id": 3164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987586.748, "dur": 7.393,
    "args": {
      "External id": 3165,"Record function id": 0, "Ev Idx": 2652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987589.572, "dur": 3.403,
    "args": {
      "External id": 3166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987590.718, "dur": 1.710,
    "args": {
      "External id": 3167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987591.584, "dur": 0.559,
    "args": {
      "External id": 3168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987599.136, "dur": 4.511,
    "args": {
      "External id": 3169,"Record function id": 0, "Ev Idx": 2656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987600.726, "dur": 2.131,
    "args": {
      "External id": 3170,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987601.494, "dur": 0.903,
    "args": {
      "External id": 3171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987601.867, "dur": 0.407,
    "args": {
      "External id": 3172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987606.957, "dur": 4.299,
    "args": {
      "External id": 3173,"Record function id": 0, "Ev Idx": 2660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987608.363, "dur": 2.110,
    "args": {
      "External id": 3174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987608.959, "dur": 1.102,
    "args": {
      "External id": 3175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987609.547, "dur": 0.372,
    "args": {
      "External id": 3176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987614.532, "dur": 5.287,
    "args": {
      "External id": 3177,"Record function id": 0, "Ev Idx": 2664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987615.878, "dur": 3.156,
    "args": {
      "External id": 3178,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987616.462, "dur": 2.133,
    "args": {
      "External id": 3179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987618.041, "dur": 0.421,
    "args": {
      "External id": 3180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987623.175, "dur": 4.187,
    "args": {
      "External id": 3181,"Record function id": 0, "Ev Idx": 2668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987624.591, "dur": 2.073,
    "args": {
      "External id": 3182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987625.322, "dur": 0.930,
    "args": {
      "External id": 3183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987625.787, "dur": 0.327,
    "args": {
      "External id": 3184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987630.678, "dur": 5.252,
    "args": {
      "External id": 3185,"Record function id": 0, "Ev Idx": 2672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987632.249, "dur": 2.860,
    "args": {
      "External id": 3186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987633.114, "dur": 0.815,
    "args": {
      "External id": 3187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987633.491, "dur": 0.310,
    "args": {
      "External id": 3188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987639.300, "dur": 4.518,
    "args": {
      "External id": 3189,"Record function id": 0, "Ev Idx": 2676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987640.872, "dur": 2.203,
    "args": {
      "External id": 3190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987641.728, "dur": 0.861,
    "args": {
      "External id": 3191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987642.140, "dur": 0.318,
    "args": {
      "External id": 3192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987647.141, "dur": 3.941,
    "args": {
      "External id": 3193,"Record function id": 0, "Ev Idx": 2680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987648.498, "dur": 1.875,
    "args": {
      "External id": 3194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987649.145, "dur": 0.821,
    "args": {
      "External id": 3195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987649.534, "dur": 0.300,
    "args": {
      "External id": 3196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987654.667, "dur": 6.999,
    "args": {
      "External id": 3197,"Record function id": 0, "Ev Idx": 2684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987656.874, "dur": 3.485,
    "args": {
      "External id": 3198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987658.217, "dur": 1.592,
    "args": {
      "External id": 3199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987658.862, "dur": 0.625,
    "args": {
      "External id": 3200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987667.820, "dur": 4.424,
    "args": {
      "External id": 3201,"Record function id": 0, "Ev Idx": 2688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987669.398, "dur": 2.149,
    "args": {
      "External id": 3202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987670.204, "dur": 0.893,
    "args": {
      "External id": 3203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987670.611, "dur": 0.340,
    "args": {
      "External id": 3204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987675.911, "dur": 11.777,
    "args": {
      "External id": 3205,"Record function id": 0, "Ev Idx": 2692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987678.065, "dur": 8.321,
    "args": {
      "External id": 3206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987682.137, "dur": 3.694,
    "args": {
      "External id": 3207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987682.935, "dur": 2.625,
    "args": {
      "External id": 3208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987692.637, "dur": 4.668,
    "args": {
      "External id": 3209,"Record function id": 0, "Ev Idx": 2696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987694.234, "dur": 2.307,
    "args": {
      "External id": 3210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987694.995, "dur": 1.118,
    "args": {
      "External id": 3211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987695.626, "dur": 0.360,
    "args": {
      "External id": 3212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987700.847, "dur": 4.255,
    "args": {
      "External id": 3213,"Record function id": 0, "Ev Idx": 2700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987702.374, "dur": 1.956,
    "args": {
      "External id": 3214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987702.946, "dur": 0.976,
    "args": {
      "External id": 3215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987703.357, "dur": 0.405,
    "args": {
      "External id": 3216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987708.338, "dur": 4.296,
    "args": {
      "External id": 3217,"Record function id": 0, "Ev Idx": 2704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987710.010, "dur": 1.845,
    "args": {
      "External id": 3218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987710.615, "dur": 0.839,
    "args": {
      "External id": 3219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987710.988, "dur": 0.332,
    "args": {
      "External id": 3220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987715.944, "dur": 4.592,
    "args": {
      "External id": 3221,"Record function id": 0, "Ev Idx": 2708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987717.254, "dur": 2.558,
    "args": {
      "External id": 3222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987717.948, "dur": 1.446,
    "args": {
      "External id": 3223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987718.326, "dur": 0.927,
    "args": {
      "External id": 3224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987723.920, "dur": 4.640,
    "args": {
      "External id": 3225,"Record function id": 0, "Ev Idx": 2712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987725.265, "dur": 2.563,
    "args": {
      "External id": 3226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 2713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987726.050, "dur": 1.374,
    "args": {
      "External id": 3227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 2714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987726.432, "dur": 0.850,
    "args": {
      "External id": 3228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 2715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987732.060, "dur": 4.061,
    "args": {
      "External id": 3229,"Record function id": 0, "Ev Idx": 2716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987733.333, "dur": 2.066,
    "args": {
      "External id": 3230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987734.035, "dur": 0.950,
    "args": {
      "External id": 3231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987734.414, "dur": 0.430,
    "args": {
      "External id": 3232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987739.422, "dur": 4.927,
    "args": {
      "External id": 3233,"Record function id": 0, "Ev Idx": 2720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987741.173, "dur": 2.402,
    "args": {
      "External id": 3234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987741.911, "dur": 1.260,
    "args": {
      "External id": 3235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987742.472, "dur": 0.560,
    "args": {
      "External id": 3236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987748.061, "dur": 85.654,
    "args": {
      "External id": 3237,"Record function id": 0, "Ev Idx": 2724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987749.533, "dur": 80.679,
    "args": {
      "External id": 3238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987750.110, "dur": 76.819,
    "args": {
      "External id": 3239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987750.591, "dur": 73.012,
    "args": {
      "External id": 3240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987857.145, "dur": 103.399,
    "args": {
      "External id": 3241,"Record function id": 0, "Ev Idx": 2728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987865.130, "dur": 94.019,
    "args": {
      "External id": 3242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987866.824, "dur": 91.494,
    "args": {
      "External id": 3243,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987952.467, "dur": 5.601,
    "args": {
      "External id": 3244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987965.833, "dur": 6.265,
    "args": {
      "External id": 3245,"Record function id": 0, "Ev Idx": 2732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987968.260, "dur": 3.006,
    "args": {
      "External id": 3246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987969.279, "dur": 1.442,
    "args": {
      "External id": 3247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987970.079, "dur": 0.486,
    "args": {
      "External id": 3248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987976.285, "dur": 6.644,
    "args": {
      "External id": 3249,"Record function id": 0, "Ev Idx": 2736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987979.447, "dur": 2.746,
    "args": {
      "External id": 3250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 2737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987980.317, "dur": 1.396,
    "args": {
      "External id": 3251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 2738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987980.934, "dur": 0.643,
    "args": {
      "External id": 3252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 2739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987987.056, "dur": 5.275,
    "args": {
      "External id": 3253,"Record function id": 0, "Ev Idx": 2740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987989.251, "dur": 2.273,
    "args": {
      "External id": 3254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987990.196, "dur": 0.835,
    "args": {
      "External id": 3255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507987990.576, "dur": 0.329,
    "args": {
      "External id": 3256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507987995.931, "dur": 10.449,
    "args": {
      "External id": 3257,"Record function id": 0, "Ev Idx": 2744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988002.195, "dur": 3.451,
    "args": {
      "External id": 3258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988003.327, "dur": 1.235,
    "args": {
      "External id": 3259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988003.891, "dur": 0.507,
    "args": {
      "External id": 3260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988010.887, "dur": 5.819,
    "args": {
      "External id": 3261,"Record function id": 0, "Ev Idx": 2748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988012.806, "dur": 3.060,
    "args": {
      "External id": 3262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988014.141, "dur": 1.078,
    "args": {
      "External id": 3263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988014.734, "dur": 0.336,
    "args": {
      "External id": 3264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988020.404, "dur": 6.490,
    "args": {
      "External id": 3265,"Record function id": 0, "Ev Idx": 2752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988022.405, "dur": 2.675,
    "args": {
      "External id": 3266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988023.222, "dur": 1.188,
    "args": {
      "External id": 3267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988023.610, "dur": 0.433,
    "args": {
      "External id": 3268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988032.234, "dur": 7.435,
    "args": {
      "External id": 3269,"Record function id": 0, "Ev Idx": 2756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988034.757, "dur": 3.810,
    "args": {
      "External id": 3270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988036.173, "dur": 1.589,
    "args": {
      "External id": 3271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988036.817, "dur": 0.677,
    "args": {
      "External id": 3272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988045.044, "dur": 6.704,
    "args": {
      "External id": 3273,"Record function id": 0, "Ev Idx": 2760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988048.527, "dur": 2.501,
    "args": {
      "External id": 3274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988049.375, "dur": 1.005,
    "args": {
      "External id": 3275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988049.799, "dur": 0.404,
    "args": {
      "External id": 3276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988055.435, "dur": 5.302,
    "args": {
      "External id": 3277,"Record function id": 0, "Ev Idx": 2764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988057.652, "dur": 2.237,
    "args": {
      "External id": 3278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988058.381, "dur": 1.024,
    "args": {
      "External id": 3279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988058.752, "dur": 0.521,
    "args": {
      "External id": 3280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988064.228, "dur": 5.565,
    "args": {
      "External id": 3281,"Record function id": 0, "Ev Idx": 2768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988066.389, "dur": 2.700,
    "args": {
      "External id": 3282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988067.377, "dur": 1.234,
    "args": {
      "External id": 3283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988068.166, "dur": 0.309,
    "args": {
      "External id": 3284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988073.379, "dur": 6.382,
    "args": {
      "External id": 3285,"Record function id": 0, "Ev Idx": 2772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988075.621, "dur": 3.306,
    "args": {
      "External id": 3286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988076.471, "dur": 1.962,
    "args": {
      "External id": 3287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988077.943, "dur": 0.350,
    "args": {
      "External id": 3288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988088.470, "dur": 5.773,
    "args": {
      "External id": 3289,"Record function id": 0, "Ev Idx": 2776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988090.382, "dur": 3.145,
    "args": {
      "External id": 3290,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988091.500, "dur": 1.506,
    "args": {
      "External id": 3291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988092.285, "dur": 0.591,
    "args": {
      "External id": 3292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988099.265, "dur": 9.795,
    "args": {
      "External id": 3293,"Record function id": 0, "Ev Idx": 2780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988102.351, "dur": 4.595,
    "args": {
      "External id": 3294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988104.481, "dur": 1.837,
    "args": {
      "External id": 3295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988105.462, "dur": 0.606,
    "args": {
      "External id": 3296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988115.034, "dur": 9.010,
    "args": {
      "External id": 3297,"Record function id": 0, "Ev Idx": 2784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988118.614, "dur": 4.037,
    "args": {
      "External id": 3298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988120.114, "dur": 1.824,
    "args": {
      "External id": 3299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988121.067, "dur": 0.624,
    "args": {
      "External id": 3300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988129.757, "dur": 9.607,
    "args": {
      "External id": 3301,"Record function id": 0, "Ev Idx": 2788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988133.467, "dur": 4.547,
    "args": {
      "External id": 3302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988135.191, "dur": 1.942,
    "args": {
      "External id": 3303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988136.290, "dur": 0.576,
    "args": {
      "External id": 3304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988144.609, "dur": 8.743,
    "args": {
      "External id": 3305,"Record function id": 0, "Ev Idx": 2792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988148.107, "dur": 4.012,
    "args": {
      "External id": 3306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988149.557, "dur": 1.652,
    "args": {
      "External id": 3307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988150.324, "dur": 0.618,
    "args": {
      "External id": 3308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988158.623, "dur": 8.079,
    "args": {
      "External id": 3309,"Record function id": 0, "Ev Idx": 2796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988161.359, "dur": 4.359,
    "args": {
      "External id": 3310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988162.470, "dur": 2.697,
    "args": {
      "External id": 3311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988163.217, "dur": 1.736,
    "args": {
      "External id": 3312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988170.532, "dur": 5.750,
    "args": {
      "External id": 3313,"Record function id": 0, "Ev Idx": 2800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988172.231, "dur": 3.201,
    "args": {
      "External id": 3314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988173.177, "dur": 1.755,
    "args": {
      "External id": 3315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988173.985, "dur": 0.815,
    "args": {
      "External id": 3316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988181.651, "dur": 8.071,
    "args": {
      "External id": 3317,"Record function id": 0, "Ev Idx": 2804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988184.486, "dur": 3.999,
    "args": {
      "External id": 3318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988186.114, "dur": 1.586,
    "args": {
      "External id": 3319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988186.814, "dur": 0.589,
    "args": {
      "External id": 3320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988195.599, "dur": 79.324,
    "args": {
      "External id": 3321,"Record function id": 0, "Ev Idx": 2808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988198.234, "dur": 74.911,
    "args": {
      "External id": 3322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988199.690, "dur": 72.429,
    "args": {
      "External id": 3323,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988200.372, "dur": 70.071,
    "args": {
      "External id": 3324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988286.669, "dur": 6.360,
    "args": {
      "External id": 3325,"Record function id": 0, "Ev Idx": 2812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988289.164, "dur": 3.078,
    "args": {
      "External id": 3326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988290.316, "dur": 1.375,
    "args": {
      "External id": 3327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988290.921, "dur": 0.643,
    "args": {
      "External id": 3328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988296.585, "dur": 5.667,
    "args": {
      "External id": 3329,"Record function id": 0, "Ev Idx": 2816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988298.219, "dur": 3.158,
    "args": {
      "External id": 3330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988299.455, "dur": 1.410,
    "args": {
      "External id": 3331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988300.052, "dur": 0.684,
    "args": {
      "External id": 3332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988305.615, "dur": 5.342,
    "args": {
      "External id": 3333,"Record function id": 0, "Ev Idx": 2820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988307.445, "dur": 2.769,
    "args": {
      "External id": 3334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988308.587, "dur": 1.132,
    "args": {
      "External id": 3335,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988309.128, "dur": 0.442,
    "args": {
      "External id": 3336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988314.477, "dur": 5.127,
    "args": {
      "External id": 3337,"Record function id": 0, "Ev Idx": 2824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988316.410, "dur": 2.330,
    "args": {
      "External id": 3338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988317.201, "dur": 1.119,
    "args": {
      "External id": 3339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988317.757, "dur": 0.432,
    "args": {
      "External id": 3340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988322.718, "dur": 5.506,
    "args": {
      "External id": 3341,"Record function id": 0, "Ev Idx": 2828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988324.605, "dur": 2.753,
    "args": {
      "External id": 3342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988325.580, "dur": 1.253,
    "args": {
      "External id": 3343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988326.328, "dur": 0.373,
    "args": {
      "External id": 3344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988331.489, "dur": 12.414,
    "args": {
      "External id": 3345,"Record function id": 0, "Ev Idx": 2832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988338.608, "dur": 4.043,
    "args": {
      "External id": 3346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988340.378, "dur": 1.443,
    "args": {
      "External id": 3347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988341.037, "dur": 0.512,
    "args": {
      "External id": 3348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988349.168, "dur": 8.432,
    "args": {
      "External id": 3349,"Record function id": 0, "Ev Idx": 2836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988351.983, "dur": 4.225,
    "args": {
      "External id": 3350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988353.648, "dur": 1.747,
    "args": {
      "External id": 3351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988354.375, "dur": 0.755,
    "args": {
      "External id": 3352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988362.502, "dur": 7.907,
    "args": {
      "External id": 3353,"Record function id": 0, "Ev Idx": 2840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988365.522, "dur": 3.779,
    "args": {
      "External id": 3354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988366.867, "dur": 1.785,
    "args": {
      "External id": 3355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988367.599, "dur": 0.758,
    "args": {
      "External id": 3356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988375.823, "dur": 8.531,
    "args": {
      "External id": 3357,"Record function id": 0, "Ev Idx": 2844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988379.286, "dur": 3.932,
    "args": {
      "External id": 3358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988380.751, "dur": 1.749,
    "args": {
      "External id": 3359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988381.753, "dur": 0.485,
    "args": {
      "External id": 3360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988389.411, "dur": 7.820,
    "args": {
      "External id": 3361,"Record function id": 0, "Ev Idx": 2848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988391.992, "dur": 4.114,
    "args": {
      "External id": 3362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988393.328, "dur": 2.133,
    "args": {
      "External id": 3363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988394.500, "dur": 0.647,
    "args": {
      "External id": 3364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988402.932, "dur": 8.100,
    "args": {
      "External id": 3365,"Record function id": 0, "Ev Idx": 2852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988405.732, "dur": 4.026,
    "args": {
      "External id": 3366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988407.278, "dur": 1.825,
    "args": {
      "External id": 3367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988408.250, "dur": 0.616,
    "args": {
      "External id": 3368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988416.155, "dur": 8.480,
    "args": {
      "External id": 3369,"Record function id": 0, "Ev Idx": 2856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988419.257, "dur": 4.154,
    "args": {
      "External id": 3370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988420.777, "dur": 1.938,
    "args": {
      "External id": 3371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988421.850, "dur": 0.614,
    "args": {
      "External id": 3372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988429.466, "dur": 8.220,
    "args": {
      "External id": 3373,"Record function id": 0, "Ev Idx": 2860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988431.993, "dur": 4.723,
    "args": {
      "External id": 3374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988433.221, "dur": 2.967,
    "args": {
      "External id": 3375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988433.857, "dur": 2.088,
    "args": {
      "External id": 3376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988441.536, "dur": 4.970,
    "args": {
      "External id": 3377,"Record function id": 0, "Ev Idx": 2864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988443.315, "dur": 2.435,
    "args": {
      "External id": 3378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988444.247, "dur": 1.034,
    "args": {
      "External id": 3379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988444.647, "dur": 0.498,
    "args": {
      "External id": 3380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988449.986, "dur": 5.635,
    "args": {
      "External id": 3381,"Record function id": 0, "Ev Idx": 2868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988452.120, "dur": 2.771,
    "args": {
      "External id": 3382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988453.111, "dur": 1.094,
    "args": {
      "External id": 3383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988453.517, "dur": 0.559,
    "args": {
      "External id": 3384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988460.970, "dur": 8.358,
    "args": {
      "External id": 3385,"Record function id": 0, "Ev Idx": 2872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988464.575, "dur": 3.567,
    "args": {
      "External id": 3386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988465.846, "dur": 1.637,
    "args": {
      "External id": 3387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988466.575, "dur": 0.664,
    "args": {
      "External id": 3388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988474.812, "dur": 78.799,
    "args": {
      "External id": 3389,"Record function id": 0, "Ev Idx": 2876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988478.053, "dur": 73.726,
    "args": {
      "External id": 3390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988479.272, "dur": 71.454,
    "args": {
      "External id": 3391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988480.124, "dur": 65.167,
    "args": {
      "External id": 3392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988567.462, "dur": 13.641,
    "args": {
      "External id": 3393,"Record function id": 0, "Ev Idx": 2880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988574.457, "dur": 5.291,
    "args": {
      "External id": 3394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988576.417, "dur": 2.443,
    "args": {
      "External id": 3395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988577.747, "dur": 0.748,
    "args": {
      "External id": 3396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988586.915, "dur": 8.763,
    "args": {
      "External id": 3397,"Record function id": 0, "Ev Idx": 2884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988589.974, "dur": 4.499,
    "args": {
      "External id": 3398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988591.746, "dur": 2.039,
    "args": {
      "External id": 3399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988592.978, "dur": 0.550,
    "args": {
      "External id": 3400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988601.135, "dur": 8.096,
    "args": {
      "External id": 3401,"Record function id": 0, "Ev Idx": 2888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988604.207, "dur": 3.640,
    "args": {
      "External id": 3402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988605.434, "dur": 1.723,
    "args": {
      "External id": 3403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988606.255, "dur": 0.640,
    "args": {
      "External id": 3404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988614.414, "dur": 8.606,
    "args": {
      "External id": 3405,"Record function id": 0, "Ev Idx": 2892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988617.166, "dur": 4.930,
    "args": {
      "External id": 3406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988618.500, "dur": 2.995,
    "args": {
      "External id": 3407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988619.178, "dur": 0.596,
    "args": {
      "External id": 3408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988626.678, "dur": 87.193,
    "args": {
      "External id": 3409,"Record function id": 0, "Ev Idx": 2896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988705.656, "dur": 6.286,
    "args": {
      "External id": 3410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988707.697, "dur": 3.379,
    "args": {
      "External id": 3411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988708.512, "dur": 2.221,
    "args": {
      "External id": 3412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988720.043, "dur": 8.963,
    "args": {
      "External id": 3413,"Record function id": 0, "Ev Idx": 2900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988723.761, "dur": 4.127,
    "args": {
      "External id": 3414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988725.420, "dur": 1.764,
    "args": {
      "External id": 3415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988726.223, "dur": 0.714,
    "args": {
      "External id": 3416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988734.678, "dur": 8.127,
    "args": {
      "External id": 3417,"Record function id": 0, "Ev Idx": 2904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988737.400, "dur": 4.221,
    "args": {
      "External id": 3418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988738.965, "dur": 1.966,
    "args": {
      "External id": 3419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988739.939, "dur": 0.696,
    "args": {
      "External id": 3420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988748.436, "dur": 8.253,
    "args": {
      "External id": 3421,"Record function id": 0, "Ev Idx": 2908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988751.089, "dur": 4.436,
    "args": {
      "External id": 3422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988752.848, "dur": 2.029,
    "args": {
      "External id": 3423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988753.764, "dur": 0.766,
    "args": {
      "External id": 3424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988762.660, "dur": 7.630,
    "args": {
      "External id": 3425,"Record function id": 0, "Ev Idx": 2912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988764.800, "dur": 4.215,
    "args": {
      "External id": 3426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988766.400, "dur": 1.966,
    "args": {
      "External id": 3427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988767.431, "dur": 0.642,
    "args": {
      "External id": 3428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988775.814, "dur": 8.908,
    "args": {
      "External id": 3429,"Record function id": 0, "Ev Idx": 2916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988779.639, "dur": 3.842,
    "args": {
      "External id": 3430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988781.244, "dur": 1.376,
    "args": {
      "External id": 3431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988781.852, "dur": 0.504,
    "args": {
      "External id": 3432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988789.985, "dur": 7.697,
    "args": {
      "External id": 3433,"Record function id": 0, "Ev Idx": 2920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988792.335, "dur": 3.905,
    "args": {
      "External id": 3434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988793.720, "dur": 1.657,
    "args": {
      "External id": 3435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988794.502, "dur": 0.602,
    "args": {
      "External id": 3436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988803.084, "dur": 9.832,
    "args": {
      "External id": 3437,"Record function id": 0, "Ev Idx": 2924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988807.565, "dur": 4.152,
    "args": {
      "External id": 3438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988809.266, "dur": 1.809,
    "args": {
      "External id": 3439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988810.276, "dur": 0.533,
    "args": {
      "External id": 3440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988818.647, "dur": 7.761,
    "args": {
      "External id": 3441,"Record function id": 0, "Ev Idx": 2928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988821.216, "dur": 3.909,
    "args": {
      "External id": 3442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988822.839, "dur": 1.650,
    "args": {
      "External id": 3443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988823.553, "dur": 0.670,
    "args": {
      "External id": 3444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988831.649, "dur": 8.069,
    "args": {
      "External id": 3445,"Record function id": 0, "Ev Idx": 2932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988834.829, "dur": 3.649,
    "args": {
      "External id": 3446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988836.281, "dur": 1.594,
    "args": {
      "External id": 3447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988837.063, "dur": 0.545,
    "args": {
      "External id": 3448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988845.112, "dur": 8.010,
    "args": {
      "External id": 3449,"Record function id": 0, "Ev Idx": 2936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988848.295, "dur": 3.937,
    "args": {
      "External id": 3450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988850.230, "dur": 1.430,
    "args": {
      "External id": 3451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988851.128, "dur": 0.377,
    "args": {
      "External id": 3452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988857.277, "dur": 4.412,
    "args": {
      "External id": 3453,"Record function id": 0, "Ev Idx": 2940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988858.726, "dur": 2.238,
    "args": {
      "External id": 3454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988859.496, "dur": 0.983,
    "args": {
      "External id": 3455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988859.879, "dur": 0.473,
    "args": {
      "External id": 3456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988867.970, "dur": 5.244,
    "args": {
      "External id": 3457,"Record function id": 0, "Ev Idx": 2944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988869.524, "dur": 2.971,
    "args": {
      "External id": 3458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988870.303, "dur": 1.720,
    "args": {
      "External id": 3459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988870.868, "dur": 1.012,
    "args": {
      "External id": 3460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988876.865, "dur": 4.476,
    "args": {
      "External id": 3461,"Record function id": 0, "Ev Idx": 2948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988878.341, "dur": 2.285,
    "args": {
      "External id": 3462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988879.194, "dur": 0.964,
    "args": {
      "External id": 3463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988879.520, "dur": 0.514,
    "args": {
      "External id": 3464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988885.047, "dur": 4.566,
    "args": {
      "External id": 3465,"Record function id": 0, "Ev Idx": 2952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988886.514, "dur": 2.387,
    "args": {
      "External id": 3466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988887.409, "dur": 1.025,
    "args": {
      "External id": 3467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988887.774, "dur": 0.490,
    "args": {
      "External id": 3468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988893.954, "dur": 8.021,
    "args": {
      "External id": 3469,"Record function id": 0, "Ev Idx": 2956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988896.436, "dur": 4.185,
    "args": {
      "External id": 3470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988898.020, "dur": 1.929,
    "args": {
      "External id": 3471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988899.057, "dur": 0.594,
    "args": {
      "External id": 3472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 2959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988907.819, "dur": 8.009,
    "args": {
      "External id": 3473,"Record function id": 0, "Ev Idx": 2960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988911.238, "dur": 3.711,
    "args": {
      "External id": 3474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988912.378, "dur": 1.926,
    "args": {
      "External id": 3475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988913.164, "dur": 1.012,
    "args": {
      "External id": 3476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988920.326, "dur": 8.520,
    "args": {
      "External id": 3477,"Record function id": 0, "Ev Idx": 2964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988923.559, "dur": 4.096,
    "args": {
      "External id": 3478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988925.235, "dur": 1.551,
    "args": {
      "External id": 3479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988925.916, "dur": 0.587,
    "args": {
      "External id": 3480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 2967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988934.781, "dur": 5.132,
    "args": {
      "External id": 3481,"Record function id": 0, "Ev Idx": 2968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988936.629, "dur": 2.533,
    "args": {
      "External id": 3482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988937.427, "dur": 1.252,
    "args": {
      "External id": 3483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988937.842, "dur": 0.708,
    "args": {
      "External id": 3484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988944.450, "dur": 104.712,
    "args": {
      "External id": 3485,"Record function id": 0, "Ev Idx": 2972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507988946.768, "dur": 97.849,
    "args": {
      "External id": 3486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988947.952, "dur": 92.043,
    "args": {
      "External id": 3487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507988948.663, "dur": 88.608,
    "args": {
      "External id": 3488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 2975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989074.647, "dur": 18.743,
    "args": {
      "External id": 3489,"Record function id": 0, "Ev Idx": 2976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989086.042, "dur": 6.572,
    "args": {
      "External id": 3490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989088.945, "dur": 3.247,
    "args": {
      "External id": 3491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989090.655, "dur": 1.396,
    "args": {
      "External id": 3492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 2979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989098.293, "dur": 6.802,
    "args": {
      "External id": 3493,"Record function id": 0, "Ev Idx": 2980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989100.694, "dur": 3.680,
    "args": {
      "External id": 3494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989102.104, "dur": 1.869,
    "args": {
      "External id": 3495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989103.302, "dur": 0.532,
    "args": {
      "External id": 3496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 2983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989109.149, "dur": 4.887,
    "args": {
      "External id": 3497,"Record function id": 0, "Ev Idx": 2984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989110.674, "dur": 2.661,
    "args": {
      "External id": 3498,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989111.883, "dur": 1.045,
    "args": {
      "External id": 3499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989112.441, "dur": 0.353,
    "args": {
      "External id": 3500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989117.119, "dur": 4.597,
    "args": {
      "External id": 3501,"Record function id": 0, "Ev Idx": 2988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989118.648, "dur": 2.355,
    "args": {
      "External id": 3502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989119.379, "dur": 1.122,
    "args": {
      "External id": 3503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989119.960, "dur": 0.394,
    "args": {
      "External id": 3504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 2991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989125.676, "dur": 5.229,
    "args": {
      "External id": 3505,"Record function id": 0, "Ev Idx": 2992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989127.974, "dur": 2.163,
    "args": {
      "External id": 3506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989128.640, "dur": 1.091,
    "args": {
      "External id": 3507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989129.258, "dur": 0.351,
    "args": {
      "External id": 3508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 2995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989134.824, "dur": 4.294,
    "args": {
      "External id": 3509,"Record function id": 0, "Ev Idx": 2996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989136.268, "dur": 2.137,
    "args": {
      "External id": 3510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989136.911, "dur": 1.066,
    "args": {
      "External id": 3511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989137.293, "dur": 0.540,
    "args": {
      "External id": 3512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 2999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989142.299, "dur": 6.689,
    "args": {
      "External id": 3513,"Record function id": 0, "Ev Idx": 3000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989143.940, "dur": 4.273,
    "args": {
      "External id": 3514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989144.981, "dur": 2.135,
    "args": {
      "External id": 3515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989145.354, "dur": 1.611,
    "args": {
      "External id": 3516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989152.186, "dur": 4.428,
    "args": {
      "External id": 3517,"Record function id": 0, "Ev Idx": 3004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989153.677, "dur": 2.145,
    "args": {
      "External id": 3518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989154.564, "dur": 0.814,
    "args": {
      "External id": 3519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989154.957, "dur": 0.298,
    "args": {
      "External id": 3520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989159.815, "dur": 4.409,
    "args": {
      "External id": 3521,"Record function id": 0, "Ev Idx": 3008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989161.416, "dur": 2.046,
    "args": {
      "External id": 3522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989162.084, "dur": 1.013,
    "args": {
      "External id": 3523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989162.457, "dur": 0.503,
    "args": {
      "External id": 3524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989167.601, "dur": 4.556,
    "args": {
      "External id": 3525,"Record function id": 0, "Ev Idx": 3012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989169.307, "dur": 2.097,
    "args": {
      "External id": 3526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989170.074, "dur": 0.944,
    "args": {
      "External id": 3527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989170.570, "dur": 0.320,
    "args": {
      "External id": 3528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989175.495, "dur": 4.483,
    "args": {
      "External id": 3529,"Record function id": 0, "Ev Idx": 3016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989177.117, "dur": 2.108,
    "args": {
      "External id": 3530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989177.807, "dur": 1.022,
    "args": {
      "External id": 3531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989178.377, "dur": 0.320,
    "args": {
      "External id": 3532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989183.269, "dur": 4.185,
    "args": {
      "External id": 3533,"Record function id": 0, "Ev Idx": 3020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989184.859, "dur": 1.832,
    "args": {
      "External id": 3534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989185.467, "dur": 0.825,
    "args": {
      "External id": 3535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989185.840, "dur": 0.324,
    "args": {
      "External id": 3536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989190.806, "dur": 4.163,
    "args": {
      "External id": 3537,"Record function id": 0, "Ev Idx": 3024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989192.248, "dur": 1.943,
    "args": {
      "External id": 3538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989192.926, "dur": 0.874,
    "args": {
      "External id": 3539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989193.331, "dur": 0.345,
    "args": {
      "External id": 3540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989200.139, "dur": 5.586,
    "args": {
      "External id": 3541,"Record function id": 0, "Ev Idx": 3028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989201.989, "dur": 2.976,
    "args": {
      "External id": 3542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989202.895, "dur": 1.674,
    "args": {
      "External id": 3543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989203.361, "dur": 1.068,
    "args": {
      "External id": 3544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989209.419, "dur": 4.215,
    "args": {
      "External id": 3545,"Record function id": 0, "Ev Idx": 3032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989210.853, "dur": 2.008,
    "args": {
      "External id": 3546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989211.519, "dur": 0.942,
    "args": {
      "External id": 3547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989211.889, "dur": 0.445,
    "args": {
      "External id": 3548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989217.423, "dur": 4.154,
    "args": {
      "External id": 3549,"Record function id": 0, "Ev Idx": 3036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989219.034, "dur": 1.823,
    "args": {
      "External id": 3550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989219.578, "dur": 0.950,
    "args": {
      "External id": 3551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989219.936, "dur": 0.457,
    "args": {
      "External id": 3552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989224.761, "dur": 4.048,
    "args": {
      "External id": 3553,"Record function id": 0, "Ev Idx": 3040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989225.938, "dur": 2.163,
    "args": {
      "External id": 3554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989226.771, "dur": 0.981,
    "args": {
      "External id": 3555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989227.245, "dur": 0.378,
    "args": {
      "External id": 3556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989232.027, "dur": 4.939,
    "args": {
      "External id": 3557,"Record function id": 0, "Ev Idx": 3044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989233.358, "dur": 2.840,
    "args": {
      "External id": 3558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989234.023, "dur": 1.792,
    "args": {
      "External id": 3559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989234.395, "dur": 1.285,
    "args": {
      "External id": 3560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989241.212, "dur": 4.637,
    "args": {
      "External id": 3561,"Record function id": 0, "Ev Idx": 3048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989242.671, "dur": 2.469,
    "args": {
      "External id": 3562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989243.319, "dur": 1.445,
    "args": {
      "External id": 3563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989243.721, "dur": 0.909,
    "args": {
      "External id": 3564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989249.022, "dur": 4.752,
    "args": {
      "External id": 3565,"Record function id": 0, "Ev Idx": 3052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989250.908, "dur": 2.143,
    "args": {
      "External id": 3566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989251.541, "dur": 1.119,
    "args": {
      "External id": 3567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989252.056, "dur": 0.472,
    "args": {
      "External id": 3568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989257.876, "dur": 64.207,
    "args": {
      "External id": 3569,"Record function id": 0, "Ev Idx": 3056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989259.397, "dur": 61.329,
    "args": {
      "External id": 3570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989260.073, "dur": 59.896,
    "args": {
      "External id": 3571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989260.445, "dur": 55.887,
    "args": {
      "External id": 3572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989328.966, "dur": 7.674,
    "args": {
      "External id": 3573,"Record function id": 0, "Ev Idx": 3060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989333.433, "dur": 2.390,
    "args": {
      "External id": 3574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989334.233, "dur": 1.143,
    "args": {
      "External id": 3575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989334.670, "dur": 0.580,
    "args": {
      "External id": 3576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989339.989, "dur": 4.264,
    "args": {
      "External id": 3577,"Record function id": 0, "Ev Idx": 3064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989341.540, "dur": 1.940,
    "args": {
      "External id": 3578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 3065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989342.155, "dur": 0.885,
    "args": {
      "External id": 3579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 3066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989342.515, "dur": 0.367,
    "args": {
      "External id": 3580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 3067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989347.442, "dur": 5.954,
    "args": {
      "External id": 3581,"Record function id": 0, "Ev Idx": 3068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989349.021, "dur": 3.541,
    "args": {
      "External id": 3582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989349.905, "dur": 2.020,
    "args": {
      "External id": 3583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989350.529, "dur": 1.243,
    "args": {
      "External id": 3584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989573.394, "dur": 68.756,
    "args": {
      "External id": 3585,"Record function id": 0, "Ev Idx": 3072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989588.724, "dur": 52.058,
    "args": {
      "External id": 3586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989590.067, "dur": 49.503,
    "args": {
      "External id": 3587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989634.938, "dur": 4.350,
    "args": {
      "External id": 3588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989646.402, "dur": 4.953,
    "args": {
      "External id": 3589,"Record function id": 0, "Ev Idx": 3076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989648.192, "dur": 2.382,
    "args": {
      "External id": 3590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989649.135, "dur": 1.049,
    "args": {
      "External id": 3591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989649.580, "dur": 0.475,
    "args": {
      "External id": 3592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989655.274, "dur": 4.545,
    "args": {
      "External id": 3593,"Record function id": 0, "Ev Idx": 3080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989656.808, "dur": 2.278,
    "args": {
      "External id": 3594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989657.627, "dur": 1.067,
    "args": {
      "External id": 3595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989658.236, "dur": 0.323,
    "args": {
      "External id": 3596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989662.978, "dur": 4.452,
    "args": {
      "External id": 3597,"Record function id": 0, "Ev Idx": 3084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989664.526, "dur": 2.225,
    "args": {
      "External id": 3598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989665.231, "dur": 1.104,
    "args": {
      "External id": 3599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989665.778, "dur": 0.427,
    "args": {
      "External id": 3600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989672.792, "dur": 4.070,
    "args": {
      "External id": 3601,"Record function id": 0, "Ev Idx": 3088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989674.279, "dur": 1.833,
    "args": {
      "External id": 3602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989674.853, "dur": 0.863,
    "args": {
      "External id": 3603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989675.224, "dur": 0.343,
    "args": {
      "External id": 3604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989680.706, "dur": 4.325,
    "args": {
      "External id": 3605,"Record function id": 0, "Ev Idx": 3092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989682.175, "dur": 2.123,
    "args": {
      "External id": 3606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989682.793, "dur": 1.083,
    "args": {
      "External id": 3607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989683.166, "dur": 0.582,
    "args": {
      "External id": 3608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507989688.421, "dur": 192.691,
    "args": {
      "External id": 3609,"Record function id": 0, "Sequence number": 4690, "Fwd thread id": 1, "Ev Idx": 3096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507989690.848, "dur": 146.955,
    "args": {
      "External id": 3610,"Sequence number": 4690, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 3097
    }
  },
  {
    "ph": "f", "id": 8, "pid": 14565, "tid": 14640, "ts": 1290507989690.848,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290507989759.310, "dur": 66.505,
    "args": {
      "External id": 3611,"Record function id": 0, "scope": "user_scope", "Ev Idx": 3098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290507989849.238, "dur": 28.970,
    "args": {
      "External id": 3612,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 3099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989886.504, "dur": 5.825,
    "args": {
      "External id": 3613,"Record function id": 0, "Ev Idx": 3100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507989888.346, "dur": 2.955,
    "args": {
      "External id": 3614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 3101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989889.485, "dur": 1.477,
    "args": {
      "External id": 3615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 3102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507989890.052, "dur": 0.755,
    "args": {
      "External id": 3616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 3103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507992638.047, "dur": 2061.622,
    "args": {
      "External id": 3617,"Record function id": 0, "Sequence number": 4693, "Fwd thread id": 1, "Ev Idx": 3104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507992640.788, "dur": 1953.658,
    "args": {
      "External id": 3618,"Sequence number": 4693, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 3105
    }
  },
  {
    "ph": "f", "id": 9, "pid": 14565, "tid": 14640, "ts": 1290507992640.788,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290507992999.236, "dur": 1567.535,
    "args": {
      "External id": 3619,"Record function id": 0, "scope": "user_scope", "Ev Idx": 3106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290507993614.823, "dur": 33.678,
    "args": {
      "External id": 3620,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 3107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994719.827, "dur": 10.897,
    "args": {
      "External id": 3621,"Record function id": 0, "Ev Idx": 3108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994723.019, "dur": 6.380,
    "args": {
      "External id": 3622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994725.915, "dur": 2.919,
    "args": {
      "External id": 3623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994726.756, "dur": 1.883,
    "args": {
      "External id": 3624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994734.783, "dur": 4.507,
    "args": {
      "External id": 3625,"Record function id": 0, "Ev Idx": 3112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994736.156, "dur": 2.375,
    "args": {
      "External id": 3626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994737.018, "dur": 1.196,
    "args": {
      "External id": 3627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994737.576, "dur": 0.495,
    "args": {
      "External id": 3628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994742.766, "dur": 4.364,
    "args": {
      "External id": 3629,"Record function id": 0, "Ev Idx": 3116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994744.044, "dur": 2.267,
    "args": {
      "External id": 3630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994744.785, "dur": 1.210,
    "args": {
      "External id": 3631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994745.394, "dur": 0.475,
    "args": {
      "External id": 3632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994750.637, "dur": 4.212,
    "args": {
      "External id": 3633,"Record function id": 0, "Ev Idx": 3120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994751.932, "dur": 2.126,
    "args": {
      "External id": 3634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994752.679, "dur": 1.053,
    "args": {
      "External id": 3635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994753.134, "dur": 0.449,
    "args": {
      "External id": 3636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994758.133, "dur": 4.340,
    "args": {
      "External id": 3637,"Record function id": 0, "Ev Idx": 3124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994759.511, "dur": 2.216,
    "args": {
      "External id": 3638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994760.234, "dur": 1.164,
    "args": {
      "External id": 3639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994760.836, "dur": 0.415,
    "args": {
      "External id": 3640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994765.731, "dur": 4.291,
    "args": {
      "External id": 3641,"Record function id": 0, "Ev Idx": 3128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994767.045, "dur": 2.247,
    "args": {
      "External id": 3642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994767.806, "dur": 1.181,
    "args": {
      "External id": 3643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994768.299, "dur": 0.555,
    "args": {
      "External id": 3644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994773.237, "dur": 4.087,
    "args": {
      "External id": 3645,"Record function id": 0, "Ev Idx": 3132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994774.681, "dur": 1.919,
    "args": {
      "External id": 3646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994775.275, "dur": 1.012,
    "args": {
      "External id": 3647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994775.786, "dur": 0.361,
    "args": {
      "External id": 3648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994780.517, "dur": 4.132,
    "args": {
      "External id": 3649,"Record function id": 0, "Ev Idx": 3136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994781.891, "dur": 2.059,
    "args": {
      "External id": 3650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994782.640, "dur": 1.006,
    "args": {
      "External id": 3651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994783.166, "dur": 0.350,
    "args": {
      "External id": 3652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994787.877, "dur": 3.836,
    "args": {
      "External id": 3653,"Record function id": 0, "Ev Idx": 3140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994789.090, "dur": 1.903,
    "args": {
      "External id": 3654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994789.783, "dur": 0.904,
    "args": {
      "External id": 3655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994790.143, "dur": 0.404,
    "args": {
      "External id": 3656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994794.895, "dur": 4.442,
    "args": {
      "External id": 3657,"Record function id": 0, "Ev Idx": 3144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994796.079, "dur": 2.456,
    "args": {
      "External id": 3658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994796.965, "dur": 0.918,
    "args": {
      "External id": 3659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994797.447, "dur": 0.302,
    "args": {
      "External id": 3660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994802.612, "dur": 4.412,
    "args": {
      "External id": 3661,"Record function id": 0, "Ev Idx": 3148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994803.944, "dur": 2.311,
    "args": {
      "External id": 3662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994804.848, "dur": 0.993,
    "args": {
      "External id": 3663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994805.254, "dur": 0.450,
    "args": {
      "External id": 3664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994810.284, "dur": 4.069,
    "args": {
      "External id": 3665,"Record function id": 0, "Ev Idx": 3152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994811.672, "dur": 1.930,
    "args": {
      "External id": 3666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994812.389, "dur": 0.910,
    "args": {
      "External id": 3667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994812.768, "dur": 0.392,
    "args": {
      "External id": 3668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994817.680, "dur": 3.786,
    "args": {
      "External id": 3669,"Record function id": 0, "Ev Idx": 3156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994818.944, "dur": 1.778,
    "args": {
      "External id": 3670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994819.551, "dur": 0.872,
    "args": {
      "External id": 3671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994819.912, "dur": 0.369,
    "args": {
      "External id": 3672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994824.658, "dur": 4.014,
    "args": {
      "External id": 3673,"Record function id": 0, "Ev Idx": 3160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994826.147, "dur": 1.799,
    "args": {
      "External id": 3674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994826.740, "dur": 0.895,
    "args": {
      "External id": 3675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994827.116, "dur": 0.378,
    "args": {
      "External id": 3676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994831.884, "dur": 3.963,
    "args": {
      "External id": 3677,"Record function id": 0, "Ev Idx": 3164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994833.299, "dur": 1.825,
    "args": {
      "External id": 3678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994833.872, "dur": 0.953,
    "args": {
      "External id": 3679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994834.372, "dur": 0.319,
    "args": {
      "External id": 3680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994839.086, "dur": 3.812,
    "args": {
      "External id": 3681,"Record function id": 0, "Ev Idx": 3168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994840.347, "dur": 1.799,
    "args": {
      "External id": 3682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994840.965, "dur": 0.869,
    "args": {
      "External id": 3683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994841.343, "dur": 0.318,
    "args": {
      "External id": 3684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994846.079, "dur": 4.106,
    "args": {
      "External id": 3685,"Record function id": 0, "Ev Idx": 3172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994847.616, "dur": 1.849,
    "args": {
      "External id": 3686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994848.245, "dur": 0.913,
    "args": {
      "External id": 3687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994848.625, "dur": 0.360,
    "args": {
      "External id": 3688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994853.560, "dur": 4.626,
    "args": {
      "External id": 3689,"Record function id": 0, "Ev Idx": 3176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994854.902, "dur": 2.553,
    "args": {
      "External id": 3690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994855.493, "dur": 1.656,
    "args": {
      "External id": 3691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994855.859, "dur": 1.150,
    "args": {
      "External id": 3692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994861.461, "dur": 3.940,
    "args": {
      "External id": 3693,"Record function id": 0, "Ev Idx": 3180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994862.716, "dur": 1.907,
    "args": {
      "External id": 3694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994863.354, "dur": 0.912,
    "args": {
      "External id": 3695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994863.734, "dur": 0.393,
    "args": {
      "External id": 3696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994868.807, "dur": 4.076,
    "args": {
      "External id": 3697,"Record function id": 0, "Ev Idx": 3184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994870.234, "dur": 1.922,
    "args": {
      "External id": 3698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994871.001, "dur": 0.853,
    "args": {
      "External id": 3699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994871.369, "dur": 0.342,
    "args": {
      "External id": 3700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994876.144, "dur": 4.510,
    "args": {
      "External id": 3701,"Record function id": 0, "Ev Idx": 3188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994877.474, "dur": 2.417,
    "args": {
      "External id": 3702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994878.187, "dur": 1.217,
    "args": {
      "External id": 3703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994878.563, "dur": 0.697,
    "args": {
      "External id": 3704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994883.979, "dur": 55.318,
    "args": {
      "External id": 3705,"Record function id": 0, "Ev Idx": 3192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994885.328, "dur": 52.557,
    "args": {
      "External id": 3706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994885.994, "dur": 50.956,
    "args": {
      "External id": 3707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994886.364, "dur": 49.308,
    "args": {
      "External id": 3708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994948.601, "dur": 7.643,
    "args": {
      "External id": 3709,"Record function id": 0, "Ev Idx": 3196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994953.215, "dur": 2.315,
    "args": {
      "External id": 3710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994954.025, "dur": 1.131,
    "args": {
      "External id": 3711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994954.605, "dur": 0.396,
    "args": {
      "External id": 3712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994960.443, "dur": 4.569,
    "args": {
      "External id": 3713,"Record function id": 0, "Ev Idx": 3200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994961.922, "dur": 2.339,
    "args": {
      "External id": 3714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994962.817, "dur": 1.080,
    "args": {
      "External id": 3715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994963.419, "dur": 0.330,
    "args": {
      "External id": 3716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994968.232, "dur": 3.876,
    "args": {
      "External id": 3717,"Record function id": 0, "Ev Idx": 3204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994969.552, "dur": 1.808,
    "args": {
      "External id": 3718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994970.174, "dur": 0.901,
    "args": {
      "External id": 3719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994970.574, "dur": 0.352,
    "args": {
      "External id": 3720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994975.284, "dur": 3.878,
    "args": {
      "External id": 3721,"Record function id": 0, "Ev Idx": 3208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994976.674, "dur": 1.775,
    "args": {
      "External id": 3722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994977.280, "dur": 0.862,
    "args": {
      "External id": 3723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994977.667, "dur": 0.320,
    "args": {
      "External id": 3724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994982.355, "dur": 3.669,
    "args": {
      "External id": 3725,"Record function id": 0, "Ev Idx": 3212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994983.542, "dur": 1.785,
    "args": {
      "External id": 3726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994984.129, "dur": 0.869,
    "args": {
      "External id": 3727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994984.506, "dur": 0.345,
    "args": {
      "External id": 3728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994989.168, "dur": 3.809,
    "args": {
      "External id": 3729,"Record function id": 0, "Ev Idx": 3216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994990.385, "dur": 1.864,
    "args": {
      "External id": 3730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994991.005, "dur": 0.910,
    "args": {
      "External id": 3731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994991.404, "dur": 0.366,
    "args": {
      "External id": 3732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994996.090, "dur": 3.804,
    "args": {
      "External id": 3733,"Record function id": 0, "Ev Idx": 3220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507994997.357, "dur": 1.838,
    "args": {
      "External id": 3734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994997.944, "dur": 0.938,
    "args": {
      "External id": 3735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507994998.304, "dur": 0.363,
    "args": {
      "External id": 3736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995003.037, "dur": 4.666,
    "args": {
      "External id": 3737,"Record function id": 0, "Ev Idx": 3224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995004.199, "dur": 2.807,
    "args": {
      "External id": 3738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995004.801, "dur": 1.848,
    "args": {
      "External id": 3739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995006.123, "dur": 0.358,
    "args": {
      "External id": 3740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995010.799, "dur": 4.067,
    "args": {
      "External id": 3741,"Record function id": 0, "Ev Idx": 3228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995012.015, "dur": 2.160,
    "args": {
      "External id": 3742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995012.822, "dur": 1.046,
    "args": {
      "External id": 3743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995013.381, "dur": 0.334,
    "args": {
      "External id": 3744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995018.012, "dur": 4.090,
    "args": {
      "External id": 3745,"Record function id": 0, "Ev Idx": 3232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995019.225, "dur": 2.165,
    "args": {
      "External id": 3746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995020.164, "dur": 0.794,
    "args": {
      "External id": 3747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995020.524, "dur": 0.291,
    "args": {
      "External id": 3748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995025.256, "dur": 4.223,
    "args": {
      "External id": 3749,"Record function id": 0, "Ev Idx": 3236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995026.664, "dur": 2.064,
    "args": {
      "External id": 3750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995027.458, "dur": 0.847,
    "args": {
      "External id": 3751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995027.867, "dur": 0.297,
    "args": {
      "External id": 3752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995032.642, "dur": 3.781,
    "args": {
      "External id": 3753,"Record function id": 0, "Ev Idx": 3240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995033.869, "dur": 1.881,
    "args": {
      "External id": 3754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995034.571, "dur": 0.844,
    "args": {
      "External id": 3755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995034.949, "dur": 0.321,
    "args": {
      "External id": 3756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995039.540, "dur": 3.683,
    "args": {
      "External id": 3757,"Record function id": 0, "Ev Idx": 3244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995040.720, "dur": 1.818,
    "args": {
      "External id": 3758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995041.315, "dur": 0.914,
    "args": {
      "External id": 3759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995041.801, "dur": 0.287,
    "args": {
      "External id": 3760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995046.356, "dur": 4.556,
    "args": {
      "External id": 3761,"Record function id": 0, "Ev Idx": 3248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995048.300, "dur": 1.878,
    "args": {
      "External id": 3762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995048.932, "dur": 0.959,
    "args": {
      "External id": 3763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995049.427, "dur": 0.319,
    "args": {
      "External id": 3764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995054.089, "dur": 6.777,
    "args": {
      "External id": 3765,"Record function id": 0, "Ev Idx": 3252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995055.550, "dur": 4.615,
    "args": {
      "External id": 3766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995058.969, "dur": 0.844,
    "args": {
      "External id": 3767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995059.338, "dur": 0.328,
    "args": {
      "External id": 3768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995064.098, "dur": 4.062,
    "args": {
      "External id": 3769,"Record function id": 0, "Ev Idx": 3256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995065.516, "dur": 1.937,
    "args": {
      "External id": 3770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995066.253, "dur": 0.870,
    "args": {
      "External id": 3771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995066.656, "dur": 0.326,
    "args": {
      "External id": 3772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995071.325, "dur": 3.645,
    "args": {
      "External id": 3773,"Record function id": 0, "Ev Idx": 3260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995072.472, "dur": 1.810,
    "args": {
      "External id": 3774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995073.092, "dur": 0.878,
    "args": {
      "External id": 3775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995073.466, "dur": 0.361,
    "args": {
      "External id": 3776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995078.082, "dur": 8.083,
    "args": {
      "External id": 3777,"Record function id": 0, "Ev Idx": 3264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995079.325, "dur": 6.118,
    "args": {
      "External id": 3778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995083.988, "dur": 1.114,
    "args": {
      "External id": 3779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995084.447, "dur": 0.508,
    "args": {
      "External id": 3780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995089.839, "dur": 4.314,
    "args": {
      "External id": 3781,"Record function id": 0, "Ev Idx": 3268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995091.430, "dur": 2.021,
    "args": {
      "External id": 3782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995092.238, "dur": 0.885,
    "args": {
      "External id": 3783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995092.643, "dur": 0.333,
    "args": {
      "External id": 3784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995097.433, "dur": 53.952,
    "args": {
      "External id": 3785,"Record function id": 0, "Ev Idx": 3272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995098.648, "dur": 50.572,
    "args": {
      "External id": 3786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995099.465, "dur": 49.106,
    "args": {
      "External id": 3787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995099.917, "dur": 47.414,
    "args": {
      "External id": 3788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995164.197, "dur": 7.994,
    "args": {
      "External id": 3789,"Record function id": 0, "Ev Idx": 3276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995168.415, "dur": 2.950,
    "args": {
      "External id": 3790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995169.596, "dur": 1.198,
    "args": {
      "External id": 3791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995170.191, "dur": 0.454,
    "args": {
      "External id": 3792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995175.557, "dur": 6.784,
    "args": {
      "External id": 3793,"Record function id": 0, "Ev Idx": 3280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995179.517, "dur": 2.090,
    "args": {
      "External id": 3794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995180.350, "dur": 0.832,
    "args": {
      "External id": 3795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995180.730, "dur": 0.326,
    "args": {
      "External id": 3796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995185.562, "dur": 4.164,
    "args": {
      "External id": 3797,"Record function id": 0, "Ev Idx": 3284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995187.172, "dur": 1.839,
    "args": {
      "External id": 3798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995187.747, "dur": 0.953,
    "args": {
      "External id": 3799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995188.259, "dur": 0.308,
    "args": {
      "External id": 3800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995192.880, "dur": 3.971,
    "args": {
      "External id": 3801,"Record function id": 0, "Ev Idx": 3288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995194.428, "dur": 1.716,
    "args": {
      "External id": 3802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995195.023, "dur": 0.802,
    "args": {
      "External id": 3803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995195.372, "dur": 0.325,
    "args": {
      "External id": 3804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995199.938, "dur": 4.065,
    "args": {
      "External id": 3805,"Record function id": 0, "Ev Idx": 3292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995201.262, "dur": 2.028,
    "args": {
      "External id": 3806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995201.908, "dur": 1.071,
    "args": {
      "External id": 3807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995202.478, "dur": 0.370,
    "args": {
      "External id": 3808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995207.130, "dur": 4.240,
    "args": {
      "External id": 3809,"Record function id": 0, "Ev Idx": 3296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995208.670, "dur": 1.997,
    "args": {
      "External id": 3810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995209.362, "dur": 0.998,
    "args": {
      "External id": 3811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995209.866, "dur": 0.364,
    "args": {
      "External id": 3812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995214.434, "dur": 3.724,
    "args": {
      "External id": 3813,"Record function id": 0, "Ev Idx": 3300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995215.634, "dur": 1.804,
    "args": {
      "External id": 3814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995216.269, "dur": 0.853,
    "args": {
      "External id": 3815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995216.679, "dur": 0.315,
    "args": {
      "External id": 3816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995221.369, "dur": 3.712,
    "args": {
      "External id": 3817,"Record function id": 0, "Ev Idx": 3304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995222.574, "dur": 1.790,
    "args": {
      "External id": 3818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995223.219, "dur": 0.835,
    "args": {
      "External id": 3819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995223.598, "dur": 0.321,
    "args": {
      "External id": 3820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995228.183, "dur": 3.882,
    "args": {
      "External id": 3821,"Record function id": 0, "Ev Idx": 3308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995229.502, "dur": 1.869,
    "args": {
      "External id": 3822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995230.218, "dur": 0.842,
    "args": {
      "External id": 3823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995230.568, "dur": 0.348,
    "args": {
      "External id": 3824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995235.150, "dur": 3.970,
    "args": {
      "External id": 3825,"Record function id": 0, "Ev Idx": 3312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995236.555, "dur": 1.858,
    "args": {
      "External id": 3826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995237.231, "dur": 0.872,
    "args": {
      "External id": 3827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995237.613, "dur": 0.341,
    "args": {
      "External id": 3828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995242.266, "dur": 4.171,
    "args": {
      "External id": 3829,"Record function id": 0, "Ev Idx": 3316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995243.889, "dur": 1.849,
    "args": {
      "External id": 3830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995244.607, "dur": 0.818,
    "args": {
      "External id": 3831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995244.939, "dur": 0.344,
    "args": {
      "External id": 3832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995249.564, "dur": 4.263,
    "args": {
      "External id": 3833,"Record function id": 0, "Ev Idx": 3320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995250.776, "dur": 2.295,
    "args": {
      "External id": 3834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995251.793, "dur": 0.869,
    "args": {
      "External id": 3835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995252.170, "dur": 0.348,
    "args": {
      "External id": 3836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995256.941, "dur": 3.879,
    "args": {
      "External id": 3837,"Record function id": 0, "Ev Idx": 3324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995258.143, "dur": 1.943,
    "args": {
      "External id": 3838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995258.837, "dur": 0.836,
    "args": {
      "External id": 3839,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995259.239, "dur": 0.302,
    "args": {
      "External id": 3840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995263.920, "dur": 4.420,
    "args": {
      "External id": 3841,"Record function id": 0, "Ev Idx": 3328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995265.155, "dur": 2.450,
    "args": {
      "External id": 3842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995265.829, "dur": 1.491,
    "args": {
      "External id": 3843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995266.332, "dur": 0.820,
    "args": {
      "External id": 3844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995271.516, "dur": 3.946,
    "args": {
      "External id": 3845,"Record function id": 0, "Ev Idx": 3332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995272.902, "dur": 1.796,
    "args": {
      "External id": 3846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995273.515, "dur": 0.844,
    "args": {
      "External id": 3847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995273.876, "dur": 0.336,
    "args": {
      "External id": 3848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995278.712, "dur": 3.949,
    "args": {
      "External id": 3849,"Record function id": 0, "Ev Idx": 3336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995279.978, "dur": 1.970,
    "args": {
      "External id": 3850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995280.732, "dur": 0.902,
    "args": {
      "External id": 3851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995281.116, "dur": 0.387,
    "args": {
      "External id": 3852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995285.778, "dur": 3.867,
    "args": {
      "External id": 3853,"Record function id": 0, "Ev Idx": 3340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995287.159, "dur": 1.777,
    "args": {
      "External id": 3854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995287.773, "dur": 0.862,
    "args": {
      "External id": 3855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995288.144, "dur": 0.352,
    "args": {
      "External id": 3856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995292.772, "dur": 3.714,
    "args": {
      "External id": 3857,"Record function id": 0, "Ev Idx": 3344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995294.004, "dur": 1.775,
    "args": {
      "External id": 3858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995294.602, "dur": 0.859,
    "args": {
      "External id": 3859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995294.977, "dur": 0.355,
    "args": {
      "External id": 3860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995299.730, "dur": 4.338,
    "args": {
      "External id": 3861,"Record function id": 0, "Ev Idx": 3348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995301.050, "dur": 2.307,
    "args": {
      "External id": 3862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995301.831, "dur": 1.216,
    "args": {
      "External id": 3863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995302.211, "dur": 0.698,
    "args": {
      "External id": 3864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995307.193, "dur": 3.709,
    "args": {
      "External id": 3865,"Record function id": 0, "Ev Idx": 3352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995308.438, "dur": 1.734,
    "args": {
      "External id": 3866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995309.033, "dur": 0.828,
    "args": {
      "External id": 3867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995309.405, "dur": 0.326,
    "args": {
      "External id": 3868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995314.139, "dur": 3.987,
    "args": {
      "External id": 3869,"Record function id": 0, "Ev Idx": 3356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995315.358, "dur": 2.059,
    "args": {
      "External id": 3870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995315.963, "dur": 1.145,
    "args": {
      "External id": 3871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995316.340, "dur": 0.627,
    "args": {
      "External id": 3872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995321.311, "dur": 53.871,
    "args": {
      "External id": 3873,"Record function id": 0, "Ev Idx": 3360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995322.592, "dur": 48.276,
    "args": {
      "External id": 3874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995323.139, "dur": 46.997,
    "args": {
      "External id": 3875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995323.511, "dur": 43.379,
    "args": {
      "External id": 3876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995385.187, "dur": 7.304,
    "args": {
      "External id": 3877,"Record function id": 0, "Ev Idx": 3364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995389.321, "dur": 2.413,
    "args": {
      "External id": 3878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995390.258, "dur": 0.945,
    "args": {
      "External id": 3879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995390.682, "dur": 0.370,
    "args": {
      "External id": 3880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995395.771, "dur": 4.326,
    "args": {
      "External id": 3881,"Record function id": 0, "Ev Idx": 3368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995397.106, "dur": 2.271,
    "args": {
      "External id": 3882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995397.925, "dur": 1.022,
    "args": {
      "External id": 3883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995398.448, "dur": 0.355,
    "args": {
      "External id": 3884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995403.221, "dur": 3.913,
    "args": {
      "External id": 3885,"Record function id": 0, "Ev Idx": 3372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995404.530, "dur": 1.892,
    "args": {
      "External id": 3886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995405.213, "dur": 0.896,
    "args": {
      "External id": 3887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995405.644, "dur": 0.315,
    "args": {
      "External id": 3888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995410.198, "dur": 3.773,
    "args": {
      "External id": 3889,"Record function id": 0, "Ev Idx": 3376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995411.520, "dur": 1.735,
    "args": {
      "External id": 3890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995412.138, "dur": 0.809,
    "args": {
      "External id": 3891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995412.505, "dur": 0.314,
    "args": {
      "External id": 3892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995417.060, "dur": 3.788,
    "args": {
      "External id": 3893,"Record function id": 0, "Ev Idx": 3380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995418.331, "dur": 1.797,
    "args": {
      "External id": 3894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995418.987, "dur": 0.834,
    "args": {
      "External id": 3895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995419.373, "dur": 0.325,
    "args": {
      "External id": 3896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995423.884, "dur": 3.722,
    "args": {
      "External id": 3897,"Record function id": 0, "Ev Idx": 3384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995425.177, "dur": 1.710,
    "args": {
      "External id": 3898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995425.729, "dur": 0.841,
    "args": {
      "External id": 3899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995426.084, "dur": 0.344,
    "args": {
      "External id": 3900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995430.672, "dur": 4.508,
    "args": {
      "External id": 3901,"Record function id": 0, "Ev Idx": 3388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995432.603, "dur": 1.866,
    "args": {
      "External id": 3902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995433.230, "dur": 0.924,
    "args": {
      "External id": 3903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995433.725, "dur": 0.298,
    "args": {
      "External id": 3904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995438.231, "dur": 3.975,
    "args": {
      "External id": 3905,"Record function id": 0, "Ev Idx": 3392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995439.489, "dur": 2.005,
    "args": {
      "External id": 3906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995440.300, "dur": 0.882,
    "args": {
      "External id": 3907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995440.734, "dur": 0.323,
    "args": {
      "External id": 3908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995445.225, "dur": 4.313,
    "args": {
      "External id": 3909,"Record function id": 0, "Ev Idx": 3396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995446.580, "dur": 2.225,
    "args": {
      "External id": 3910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995447.540, "dur": 0.858,
    "args": {
      "External id": 3911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995447.926, "dur": 0.330,
    "args": {
      "External id": 3912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995452.715, "dur": 4.165,
    "args": {
      "External id": 3913,"Record function id": 0, "Ev Idx": 3400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995454.097, "dur": 2.060,
    "args": {
      "External id": 3914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995454.953, "dur": 0.789,
    "args": {
      "External id": 3915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995455.322, "dur": 0.294,
    "args": {
      "External id": 3916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995459.957, "dur": 4.010,
    "args": {
      "External id": 3917,"Record function id": 0, "Ev Idx": 3404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995461.265, "dur": 1.941,
    "args": {
      "External id": 3918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995461.868, "dur": 1.031,
    "args": {
      "External id": 3919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995462.432, "dur": 0.323,
    "args": {
      "External id": 3920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995467.052, "dur": 4.049,
    "args": {
      "External id": 3921,"Record function id": 0, "Ev Idx": 3408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995468.270, "dur": 2.079,
    "args": {
      "External id": 3922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995468.953, "dur": 0.986,
    "args": {
      "External id": 3923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995469.489, "dur": 0.321,
    "args": {
      "External id": 3924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995474.208, "dur": 5.344,
    "args": {
      "External id": 3925,"Record function id": 0, "Ev Idx": 3412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995475.498, "dur": 3.335,
    "args": {
      "External id": 3926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995476.132, "dur": 2.362,
    "args": {
      "External id": 3927,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995476.511, "dur": 1.796,
    "args": {
      "External id": 3928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995483.431, "dur": 4.140,
    "args": {
      "External id": 3929,"Record function id": 0, "Ev Idx": 3416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995484.761, "dur": 2.073,
    "args": {
      "External id": 3930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995485.465, "dur": 1.034,
    "args": {
      "External id": 3931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995486.009, "dur": 0.351,
    "args": {
      "External id": 3932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995490.840, "dur": 3.749,
    "args": {
      "External id": 3933,"Record function id": 0, "Ev Idx": 3420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995492.099, "dur": 1.773,
    "args": {
      "External id": 3934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995492.682, "dur": 0.897,
    "args": {
      "External id": 3935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995493.060, "dur": 0.391,
    "args": {
      "External id": 3936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995497.655, "dur": 3.748,
    "args": {
      "External id": 3937,"Record function id": 0, "Ev Idx": 3424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995498.945, "dur": 1.758,
    "args": {
      "External id": 3938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995499.544, "dur": 0.849,
    "args": {
      "External id": 3939,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995499.878, "dur": 0.346,
    "args": {
      "External id": 3940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995504.489, "dur": 4.209,
    "args": {
      "External id": 3941,"Record function id": 0, "Ev Idx": 3428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995505.840, "dur": 2.155,
    "args": {
      "External id": 3942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995506.456, "dur": 1.234,
    "args": {
      "External id": 3943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995506.842, "dur": 0.709,
    "args": {
      "External id": 3944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995511.828, "dur": 4.040,
    "args": {
      "External id": 3945,"Record function id": 0, "Ev Idx": 3432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995513.074, "dur": 2.044,
    "args": {
      "External id": 3946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995513.686, "dur": 1.150,
    "args": {
      "External id": 3947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995514.091, "dur": 0.610,
    "args": {
      "External id": 3948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995519.126, "dur": 4.057,
    "args": {
      "External id": 3949,"Record function id": 0, "Ev Idx": 3436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995520.443, "dur": 2.021,
    "args": {
      "External id": 3950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995521.176, "dur": 1.015,
    "args": {
      "External id": 3951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995521.507, "dur": 0.549,
    "args": {
      "External id": 3952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995526.238, "dur": 4.273,
    "args": {
      "External id": 3953,"Record function id": 0, "Ev Idx": 3440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995527.733, "dur": 2.035,
    "args": {
      "External id": 3954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995528.347, "dur": 1.124,
    "args": {
      "External id": 3955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995528.810, "dur": 0.528,
    "args": {
      "External id": 3956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995533.703, "dur": 54.271,
    "args": {
      "External id": 3957,"Record function id": 0, "Ev Idx": 3444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995535.127, "dur": 51.525,
    "args": {
      "External id": 3958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995535.805, "dur": 49.421,
    "args": {
      "External id": 3959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995536.247, "dur": 43.039,
    "args": {
      "External id": 3960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995596.803, "dur": 7.572,
    "args": {
      "External id": 3961,"Record function id": 0, "Ev Idx": 3448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995600.893, "dur": 2.716,
    "args": {
      "External id": 3962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995602.026, "dur": 1.062,
    "args": {
      "External id": 3963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995602.560, "dur": 0.348,
    "args": {
      "External id": 3964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995607.613, "dur": 4.230,
    "args": {
      "External id": 3965,"Record function id": 0, "Ev Idx": 3452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995608.917, "dur": 2.137,
    "args": {
      "External id": 3966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995609.728, "dur": 0.912,
    "args": {
      "External id": 3967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995610.183, "dur": 0.311,
    "args": {
      "External id": 3968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995614.932, "dur": 3.810,
    "args": {
      "External id": 3969,"Record function id": 0, "Ev Idx": 3456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995616.299, "dur": 1.763,
    "args": {
      "External id": 3970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995616.916, "dur": 0.831,
    "args": {
      "External id": 3971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995617.291, "dur": 0.324,
    "args": {
      "External id": 3972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995621.796, "dur": 3.929,
    "args": {
      "External id": 3973,"Record function id": 0, "Ev Idx": 3460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995623.217, "dur": 1.824,
    "args": {
      "External id": 3974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995623.854, "dur": 0.874,
    "args": {
      "External id": 3975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995624.252, "dur": 0.328,
    "args": {
      "External id": 3976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995628.774, "dur": 3.638,
    "args": {
      "External id": 3977,"Record function id": 0, "Ev Idx": 3464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995630.009, "dur": 1.720,
    "args": {
      "External id": 3978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995630.606, "dur": 0.809,
    "args": {
      "External id": 3979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995630.979, "dur": 0.312,
    "args": {
      "External id": 3980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995635.458, "dur": 3.713,
    "args": {
      "External id": 3981,"Record function id": 0, "Ev Idx": 3468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995636.792, "dur": 1.700,
    "args": {
      "External id": 3982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995637.383, "dur": 0.794,
    "args": {
      "External id": 3983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995637.733, "dur": 0.318,
    "args": {
      "External id": 3984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995642.250, "dur": 58.645,
    "args": {
      "External id": 3985,"Record function id": 0, "Ev Idx": 3472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995643.506, "dur": 56.024,
    "args": {
      "External id": 3986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995692.482, "dur": 4.521,
    "args": {
      "External id": 3987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995693.467, "dur": 3.300,
    "args": {
      "External id": 3988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995704.829, "dur": 4.429,
    "args": {
      "External id": 3989,"Record function id": 0, "Ev Idx": 3476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995706.281, "dur": 2.248,
    "args": {
      "External id": 3990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995707.071, "dur": 1.147,
    "args": {
      "External id": 3991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995707.629, "dur": 0.462,
    "args": {
      "External id": 3992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995712.528, "dur": 4.240,
    "args": {
      "External id": 3993,"Record function id": 0, "Ev Idx": 3480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995713.999, "dur": 2.077,
    "args": {
      "External id": 3994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995714.731, "dur": 1.050,
    "args": {
      "External id": 3995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995715.315, "dur": 0.339,
    "args": {
      "External id": 3996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995720.118, "dur": 3.722,
    "args": {
      "External id": 3997,"Record function id": 0, "Ev Idx": 3484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995721.383, "dur": 1.767,
    "args": {
      "External id": 3998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995722.067, "dur": 0.801,
    "args": {
      "External id": 3999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995722.430, "dur": 0.314,
    "args": {
      "External id": 4000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995727.040, "dur": 4.066,
    "args": {
      "External id": 4001,"Record function id": 0, "Ev Idx": 3488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995728.526, "dur": 1.915,
    "args": {
      "External id": 4002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995729.308, "dur": 0.832,
    "args": {
      "External id": 4003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995729.692, "dur": 0.310,
    "args": {
      "External id": 4004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995734.239, "dur": 3.848,
    "args": {
      "External id": 4005,"Record function id": 0, "Ev Idx": 3492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995735.415, "dur": 1.999,
    "args": {
      "External id": 4006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995736.102, "dur": 1.002,
    "args": {
      "External id": 4007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995736.610, "dur": 0.348,
    "args": {
      "External id": 4008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995741.221, "dur": 4.264,
    "args": {
      "External id": 4009,"Record function id": 0, "Ev Idx": 3496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995742.513, "dur": 2.269,
    "args": {
      "External id": 4010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995743.486, "dur": 0.984,
    "args": {
      "External id": 4011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995743.930, "dur": 0.338,
    "args": {
      "External id": 4012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995748.668, "dur": 3.900,
    "args": {
      "External id": 4013,"Record function id": 0, "Ev Idx": 3500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995749.994, "dur": 1.889,
    "args": {
      "External id": 4014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995750.592, "dur": 0.981,
    "args": {
      "External id": 4015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995751.128, "dur": 0.319,
    "args": {
      "External id": 4016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995755.711, "dur": 5.588,
    "args": {
      "External id": 4017,"Record function id": 0, "Ev Idx": 3504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995756.984, "dur": 3.588,
    "args": {
      "External id": 4018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995757.579, "dur": 2.635,
    "args": {
      "External id": 4019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995758.182, "dur": 1.843,
    "args": {
      "External id": 4020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995764.512, "dur": 3.832,
    "args": {
      "External id": 4021,"Record function id": 0, "Ev Idx": 3508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995765.799, "dur": 1.850,
    "args": {
      "External id": 4022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995766.434, "dur": 0.870,
    "args": {
      "External id": 4023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995766.814, "dur": 0.337,
    "args": {
      "External id": 4024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995771.643, "dur": 4.008,
    "args": {
      "External id": 4025,"Record function id": 0, "Ev Idx": 3512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995773.123, "dur": 1.832,
    "args": {
      "External id": 4026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 3513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995773.772, "dur": 0.890,
    "args": {
      "External id": 4027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 3514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995774.168, "dur": 0.352,
    "args": {
      "External id": 4028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 3515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995778.868, "dur": 4.554,
    "args": {
      "External id": 4029,"Record function id": 0, "Ev Idx": 3516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995780.479, "dur": 2.234,
    "args": {
      "External id": 4030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995781.222, "dur": 1.192,
    "args": {
      "External id": 4031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995781.585, "dur": 0.632,
    "args": {
      "External id": 4032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995786.607, "dur": 58.500,
    "args": {
      "External id": 4033,"Record function id": 0, "Ev Idx": 3520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995834.807, "dur": 8.957,
    "args": {
      "External id": 4034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995838.397, "dur": 4.808,
    "args": {
      "External id": 4035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995838.832, "dur": 4.177,
    "args": {
      "External id": 4036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995849.065, "dur": 4.578,
    "args": {
      "External id": 4037,"Record function id": 0, "Ev Idx": 3524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995850.570, "dur": 2.338,
    "args": {
      "External id": 4038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995851.449, "dur": 1.056,
    "args": {
      "External id": 4039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995851.799, "dur": 0.570,
    "args": {
      "External id": 4040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995856.913, "dur": 54.691,
    "args": {
      "External id": 4041,"Record function id": 0, "Ev Idx": 3528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995858.435, "dur": 49.387,
    "args": {
      "External id": 4042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995859.145, "dur": 47.966,
    "args": {
      "External id": 4043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995859.535, "dur": 44.812,
    "args": {
      "External id": 4044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995920.411, "dur": 7.120,
    "args": {
      "External id": 4045,"Record function id": 0, "Ev Idx": 3532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995924.719, "dur": 2.105,
    "args": {
      "External id": 4046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995925.553, "dur": 0.938,
    "args": {
      "External id": 4047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995925.963, "dur": 0.400,
    "args": {
      "External id": 4048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995930.665, "dur": 3.734,
    "args": {
      "External id": 4049,"Record function id": 0, "Ev Idx": 3536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995931.911, "dur": 1.780,
    "args": {
      "External id": 4050,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995932.557, "dur": 0.824,
    "args": {
      "External id": 4051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995932.902, "dur": 0.322,
    "args": {
      "External id": 4052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995937.491, "dur": 4.367,
    "args": {
      "External id": 4053,"Record function id": 0, "Ev Idx": 3540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995938.993, "dur": 2.165,
    "args": {
      "External id": 4054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995939.758, "dur": 1.085,
    "args": {
      "External id": 4055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995940.385, "dur": 0.328,
    "args": {
      "External id": 4056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995944.957, "dur": 3.559,
    "args": {
      "External id": 4057,"Record function id": 0, "Ev Idx": 3544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995946.125, "dur": 1.705,
    "args": {
      "External id": 4058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995946.695, "dur": 0.825,
    "args": {
      "External id": 4059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995947.060, "dur": 0.320,
    "args": {
      "External id": 4060,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995951.538, "dur": 3.716,
    "args": {
      "External id": 4061,"Record function id": 0, "Ev Idx": 3548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995952.771, "dur": 1.789,
    "args": {
      "External id": 4062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995953.428, "dur": 0.822,
    "args": {
      "External id": 4063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995953.757, "dur": 0.357,
    "args": {
      "External id": 4064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995958.375, "dur": 3.949,
    "args": {
      "External id": 4065,"Record function id": 0, "Ev Idx": 3552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995959.622, "dur": 1.994,
    "args": {
      "External id": 4066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995960.314, "dur": 0.970,
    "args": {
      "External id": 4067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995960.676, "dur": 0.460,
    "args": {
      "External id": 4068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995965.355, "dur": 3.745,
    "args": {
      "External id": 4069,"Record function id": 0, "Ev Idx": 3556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995966.580, "dur": 1.822,
    "args": {
      "External id": 4070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995967.200, "dur": 0.865,
    "args": {
      "External id": 4071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995967.553, "dur": 0.373,
    "args": {
      "External id": 4072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995972.163, "dur": 4.026,
    "args": {
      "External id": 4073,"Record function id": 0, "Ev Idx": 3560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995973.608, "dur": 1.853,
    "args": {
      "External id": 4074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995974.296, "dur": 0.853,
    "args": {
      "External id": 4075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995974.676, "dur": 0.351,
    "args": {
      "External id": 4076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995979.312, "dur": 4.165,
    "args": {
      "External id": 4077,"Record function id": 0, "Ev Idx": 3564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995980.516, "dur": 2.208,
    "args": {
      "External id": 4078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995981.218, "dur": 0.970,
    "args": {
      "External id": 4079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995981.739, "dur": 0.303,
    "args": {
      "External id": 4080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995986.570, "dur": 3.814,
    "args": {
      "External id": 4081,"Record function id": 0, "Ev Idx": 3568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995987.746, "dur": 1.912,
    "args": {
      "External id": 4082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995988.449, "dur": 0.797,
    "args": {
      "External id": 4083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995988.811, "dur": 0.300,
    "args": {
      "External id": 4084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995993.516, "dur": 3.767,
    "args": {
      "External id": 4085,"Record function id": 0, "Ev Idx": 3572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507995994.821, "dur": 1.768,
    "args": {
      "External id": 4086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995995.368, "dur": 0.901,
    "args": {
      "External id": 4087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507995995.716, "dur": 0.424,
    "args": {
      "External id": 4088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996000.337, "dur": 4.353,
    "args": {
      "External id": 4089,"Record function id": 0, "Ev Idx": 3576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996001.619, "dur": 2.348,
    "args": {
      "External id": 4090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996002.553, "dur": 1.102,
    "args": {
      "External id": 4091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996003.164, "dur": 0.364,
    "args": {
      "External id": 4092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996007.840, "dur": 4.339,
    "args": {
      "External id": 4093,"Record function id": 0, "Ev Idx": 3580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996009.064, "dur": 2.429,
    "args": {
      "External id": 4094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996009.677, "dur": 1.497,
    "args": {
      "External id": 4095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996010.101, "dur": 0.933,
    "args": {
      "External id": 4096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996233.583, "dur": 24.138,
    "args": {
      "External id": 4097,"Record function id": 0, "Ev Idx": 3584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996248.417, "dur": 8.391,
    "args": {
      "External id": 4098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996249.396, "dur": 6.870,
    "args": {
      "External id": 4099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996252.406, "dur": 3.633,
    "args": {
      "External id": 4100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996261.533, "dur": 3.885,
    "args": {
      "External id": 4101,"Record function id": 0, "Ev Idx": 3588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996262.914, "dur": 1.829,
    "args": {
      "External id": 4102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996263.533, "dur": 0.894,
    "args": {
      "External id": 4103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996263.913, "dur": 0.370,
    "args": {
      "External id": 4104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996268.663, "dur": 3.667,
    "args": {
      "External id": 4105,"Record function id": 0, "Ev Idx": 3592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996269.962, "dur": 1.690,
    "args": {
      "External id": 4106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996270.494, "dur": 0.854,
    "args": {
      "External id": 4107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996270.873, "dur": 0.347,
    "args": {
      "External id": 4108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996275.550, "dur": 3.784,
    "args": {
      "External id": 4109,"Record function id": 0, "Ev Idx": 3596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996276.929, "dur": 1.727,
    "args": {
      "External id": 4110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996277.500, "dur": 0.855,
    "args": {
      "External id": 4111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996277.870, "dur": 0.357,
    "args": {
      "External id": 4112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996282.476, "dur": 3.988,
    "args": {
      "External id": 4113,"Record function id": 0, "Ev Idx": 3600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996283.667, "dur": 2.115,
    "args": {
      "External id": 4114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996284.216, "dur": 1.264,
    "args": {
      "External id": 4115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996284.636, "dur": 0.713,
    "args": {
      "External id": 4116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996289.696, "dur": 3.832,
    "args": {
      "External id": 4117,"Record function id": 0, "Ev Idx": 3604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996290.936, "dur": 1.919,
    "args": {
      "External id": 4118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996291.500, "dur": 1.056,
    "args": {
      "External id": 4119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996291.888, "dur": 0.533,
    "args": {
      "External id": 4120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996296.805, "dur": 54.527,
    "args": {
      "External id": 4121,"Record function id": 0, "Ev Idx": 3608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996298.078, "dur": 51.918,
    "args": {
      "External id": 4122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996298.823, "dur": 50.236,
    "args": {
      "External id": 4123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996299.955, "dur": 44.874,
    "args": {
      "External id": 4124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996358.394, "dur": 8.082,
    "args": {
      "External id": 4125,"Record function id": 0, "Ev Idx": 3612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996363.406, "dur": 2.328,
    "args": {
      "External id": 4126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996364.387, "dur": 0.892,
    "args": {
      "External id": 4127,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996364.824, "dur": 0.313,
    "args": {
      "External id": 4128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996369.664, "dur": 3.712,
    "args": {
      "External id": 4129,"Record function id": 0, "Ev Idx": 3616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996370.884, "dur": 1.806,
    "args": {
      "External id": 4130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996371.513, "dur": 0.846,
    "args": {
      "External id": 4131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996371.912, "dur": 0.314,
    "args": {
      "External id": 4132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996376.495, "dur": 4.000,
    "args": {
      "External id": 4133,"Record function id": 0, "Ev Idx": 3620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996377.885, "dur": 1.898,
    "args": {
      "External id": 4134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996378.558, "dur": 0.911,
    "args": {
      "External id": 4135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996378.948, "dur": 0.350,
    "args": {
      "External id": 4136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996383.540, "dur": 3.752,
    "args": {
      "External id": 4137,"Record function id": 0, "Ev Idx": 3624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996384.750, "dur": 1.824,
    "args": {
      "External id": 4138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996385.338, "dur": 0.901,
    "args": {
      "External id": 4139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996385.737, "dur": 0.348,
    "args": {
      "External id": 4140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996390.385, "dur": 4.105,
    "args": {
      "External id": 4141,"Record function id": 0, "Ev Idx": 3628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996391.735, "dur": 2.053,
    "args": {
      "External id": 4142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996392.354, "dur": 1.031,
    "args": {
      "External id": 4143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996392.868, "dur": 0.373,
    "args": {
      "External id": 4144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996397.555, "dur": 4.046,
    "args": {
      "External id": 4145,"Record function id": 0, "Ev Idx": 3632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996398.869, "dur": 2.032,
    "args": {
      "External id": 4146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996399.573, "dur": 1.009,
    "args": {
      "External id": 4147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996399.945, "dur": 0.493,
    "args": {
      "External id": 4148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996404.664, "dur": 4.638,
    "args": {
      "External id": 4149,"Record function id": 0, "Ev Idx": 3636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996405.850, "dur": 2.745,
    "args": {
      "External id": 4150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996407.252, "dur": 1.000,
    "args": {
      "External id": 4151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996407.601, "dur": 0.512,
    "args": {
      "External id": 4152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996412.365, "dur": 4.023,
    "args": {
      "External id": 4153,"Record function id": 0, "Ev Idx": 3640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996413.718, "dur": 1.995,
    "args": {
      "External id": 4154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996414.315, "dur": 1.063,
    "args": {
      "External id": 4155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996414.684, "dur": 0.555,
    "args": {
      "External id": 4156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996419.458, "dur": 3.649,
    "args": {
      "External id": 4157,"Record function id": 0, "Ev Idx": 3644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996420.602, "dur": 1.804,
    "args": {
      "External id": 4158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996421.200, "dur": 0.872,
    "args": {
      "External id": 4159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996421.561, "dur": 0.361,
    "args": {
      "External id": 4160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996426.155, "dur": 3.904,
    "args": {
      "External id": 4161,"Record function id": 0, "Ev Idx": 3648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996427.400, "dur": 1.954,
    "args": {
      "External id": 4162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996428.101, "dur": 0.920,
    "args": {
      "External id": 4163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996428.480, "dur": 0.408,
    "args": {
      "External id": 4164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996433.076, "dur": 3.903,
    "args": {
      "External id": 4165,"Record function id": 0, "Ev Idx": 3652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996434.448, "dur": 1.856,
    "args": {
      "External id": 4166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996435.143, "dur": 0.842,
    "args": {
      "External id": 4167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996435.515, "dur": 0.345,
    "args": {
      "External id": 4168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996439.982, "dur": 4.137,
    "args": {
      "External id": 4169,"Record function id": 0, "Ev Idx": 3656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996441.206, "dur": 2.160,
    "args": {
      "External id": 4170,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996442.105, "dur": 0.825,
    "args": {
      "External id": 4171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996442.452, "dur": 0.336,
    "args": {
      "External id": 4172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996447.170, "dur": 3.801,
    "args": {
      "External id": 4173,"Record function id": 0, "Ev Idx": 3660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996448.505, "dur": 1.761,
    "args": {
      "External id": 4174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996449.102, "dur": 0.854,
    "args": {
      "External id": 4175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996449.455, "dur": 0.346,
    "args": {
      "External id": 4176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996454.651, "dur": 4.478,
    "args": {
      "External id": 4177,"Record function id": 0, "Ev Idx": 3664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996456.609, "dur": 1.812,
    "args": {
      "External id": 4178,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996457.255, "dur": 0.840,
    "args": {
      "External id": 4179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996457.637, "dur": 0.326,
    "args": {
      "External id": 4180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996462.288, "dur": 3.999,
    "args": {
      "External id": 4181,"Record function id": 0, "Ev Idx": 3668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996463.803, "dur": 1.772,
    "args": {
      "External id": 4182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996464.419, "dur": 0.843,
    "args": {
      "External id": 4183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996464.766, "dur": 0.361,
    "args": {
      "External id": 4184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996469.345, "dur": 3.827,
    "args": {
      "External id": 4185,"Record function id": 0, "Ev Idx": 3672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996470.787, "dur": 1.695,
    "args": {
      "External id": 4186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996471.355, "dur": 0.819,
    "args": {
      "External id": 4187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996471.703, "dur": 0.346,
    "args": {
      "External id": 4188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996476.219, "dur": 4.362,
    "args": {
      "External id": 4189,"Record function id": 0, "Ev Idx": 3676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996477.504, "dur": 2.383,
    "args": {
      "External id": 4190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996478.269, "dur": 1.304,
    "args": {
      "External id": 4191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996478.801, "dur": 0.646,
    "args": {
      "External id": 4192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996483.691, "dur": 4.447,
    "args": {
      "External id": 4193,"Record function id": 0, "Ev Idx": 3680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996485.023, "dur": 2.415,
    "args": {
      "External id": 4194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996485.823, "dur": 1.302,
    "args": {
      "External id": 4195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996486.175, "dur": 0.819,
    "args": {
      "External id": 4196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996491.226, "dur": 4.280,
    "args": {
      "External id": 4197,"Record function id": 0, "Ev Idx": 3684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996492.498, "dur": 2.309,
    "args": {
      "External id": 4198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996493.348, "dur": 1.136,
    "args": {
      "External id": 4199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996493.726, "dur": 0.631,
    "args": {
      "External id": 4200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996498.659, "dur": 4.087,
    "args": {
      "External id": 4201,"Record function id": 0, "Ev Idx": 3688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996500.131, "dur": 1.918,
    "args": {
      "External id": 4202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996500.895, "dur": 0.839,
    "args": {
      "External id": 4203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996501.236, "dur": 0.357,
    "args": {
      "External id": 4204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996506.434, "dur": 4.426,
    "args": {
      "External id": 4205,"Record function id": 0, "Ev Idx": 3692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996507.848, "dur": 2.239,
    "args": {
      "External id": 4206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996508.614, "dur": 1.040,
    "args": {
      "External id": 4207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996508.960, "dur": 0.549,
    "args": {
      "External id": 4208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996514.194, "dur": 51.064,
    "args": {
      "External id": 4209,"Record function id": 0, "Ev Idx": 3696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996515.467, "dur": 46.139,
    "args": {
      "External id": 4210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996516.355, "dur": 44.380,
    "args": {
      "External id": 4211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996516.726, "dur": 40.845,
    "args": {
      "External id": 4212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996573.641, "dur": 7.895,
    "args": {
      "External id": 4213,"Record function id": 0, "Ev Idx": 3700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996578.544, "dur": 2.276,
    "args": {
      "External id": 4214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996579.399, "dur": 1.129,
    "args": {
      "External id": 4215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996579.857, "dur": 0.541,
    "args": {
      "External id": 4216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996584.791, "dur": 3.705,
    "args": {
      "External id": 4217,"Record function id": 0, "Ev Idx": 3704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996586.083, "dur": 1.737,
    "args": {
      "External id": 4218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996586.698, "dur": 0.810,
    "args": {
      "External id": 4219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996587.034, "dur": 0.326,
    "args": {
      "External id": 4220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996591.543, "dur": 3.861,
    "args": {
      "External id": 4221,"Record function id": 0, "Ev Idx": 3708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996592.982, "dur": 1.741,
    "args": {
      "External id": 4222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996593.608, "dur": 0.810,
    "args": {
      "External id": 4223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996593.954, "dur": 0.327,
    "args": {
      "External id": 4224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996598.485, "dur": 3.603,
    "args": {
      "External id": 4225,"Record function id": 0, "Ev Idx": 3712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996599.699, "dur": 1.700,
    "args": {
      "External id": 4226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996600.313, "dur": 0.782,
    "args": {
      "External id": 4227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996600.643, "dur": 0.326,
    "args": {
      "External id": 4228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996605.180, "dur": 4.352,
    "args": {
      "External id": 4229,"Record function id": 0, "Ev Idx": 3716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996606.343, "dur": 2.508,
    "args": {
      "External id": 4230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996606.912, "dur": 1.597,
    "args": {
      "External id": 4231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996608.005, "dur": 0.359,
    "args": {
      "External id": 4232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996612.634, "dur": 3.964,
    "args": {
      "External id": 4233,"Record function id": 0, "Ev Idx": 3720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996614.052, "dur": 1.849,
    "args": {
      "External id": 4234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996614.633, "dur": 0.959,
    "args": {
      "External id": 4235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996614.979, "dur": 0.481,
    "args": {
      "External id": 4236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996619.648, "dur": 3.746,
    "args": {
      "External id": 4237,"Record function id": 0, "Ev Idx": 3724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996621.011, "dur": 1.687,
    "args": {
      "External id": 4238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996621.562, "dur": 0.818,
    "args": {
      "External id": 4239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996621.899, "dur": 0.354,
    "args": {
      "External id": 4240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996626.445, "dur": 3.716,
    "args": {
      "External id": 4241,"Record function id": 0, "Ev Idx": 3728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996627.695, "dur": 1.796,
    "args": {
      "External id": 4242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996628.321, "dur": 0.854,
    "args": {
      "External id": 4243,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996628.663, "dur": 0.367,
    "args": {
      "External id": 4244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996633.258, "dur": 4.216,
    "args": {
      "External id": 4245,"Record function id": 0, "Ev Idx": 3732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996634.582, "dur": 2.207,
    "args": {
      "External id": 4246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996635.480, "dur": 0.998,
    "args": {
      "External id": 4247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996636.014, "dur": 0.330,
    "args": {
      "External id": 4248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996640.569, "dur": 4.122,
    "args": {
      "External id": 4249,"Record function id": 0, "Ev Idx": 3736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996641.842, "dur": 2.121,
    "args": {
      "External id": 4250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996642.594, "dur": 0.951,
    "args": {
      "External id": 4251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996643.094, "dur": 0.305,
    "args": {
      "External id": 4252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996647.854, "dur": 4.241,
    "args": {
      "External id": 4253,"Record function id": 0, "Ev Idx": 3740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996649.195, "dur": 2.183,
    "args": {
      "External id": 4254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996649.971, "dur": 0.999,
    "args": {
      "External id": 4255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996650.543, "dur": 0.295,
    "args": {
      "External id": 4256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996655.117, "dur": 5.987,
    "args": {
      "External id": 4257,"Record function id": 0, "Ev Idx": 3744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996656.300, "dur": 4.089,
    "args": {
      "External id": 4258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996659.145, "dur": 0.899,
    "args": {
      "External id": 4259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996659.591, "dur": 0.321,
    "args": {
      "External id": 4260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996664.197, "dur": 5.295,
    "args": {
      "External id": 4261,"Record function id": 0, "Ev Idx": 3748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996666.056, "dur": 2.705,
    "args": {
      "External id": 4262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996666.661, "dur": 1.774,
    "args": {
      "External id": 4263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996667.031, "dur": 1.229,
    "args": {
      "External id": 4264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996672.663, "dur": 3.792,
    "args": {
      "External id": 4265,"Record function id": 0, "Ev Idx": 3752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996673.951, "dur": 1.819,
    "args": {
      "External id": 4266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996674.581, "dur": 0.885,
    "args": {
      "External id": 4267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996674.983, "dur": 0.348,
    "args": {
      "External id": 4268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996679.588, "dur": 3.952,
    "args": {
      "External id": 4269,"Record function id": 0, "Ev Idx": 3756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996680.977, "dur": 1.874,
    "args": {
      "External id": 4270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996681.663, "dur": 0.879,
    "args": {
      "External id": 4271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996682.057, "dur": 0.355,
    "args": {
      "External id": 4272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996686.589, "dur": 3.712,
    "args": {
      "External id": 4273,"Record function id": 0, "Ev Idx": 3760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996687.819, "dur": 1.795,
    "args": {
      "External id": 4274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996688.439, "dur": 0.860,
    "args": {
      "External id": 4275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996688.827, "dur": 0.345,
    "args": {
      "External id": 4276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996693.364, "dur": 4.935,
    "args": {
      "External id": 4277,"Record function id": 0, "Ev Idx": 3764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996694.544, "dur": 3.059,
    "args": {
      "External id": 4278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996695.961, "dur": 1.291,
    "args": {
      "External id": 4279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996696.385, "dur": 0.742,
    "args": {
      "External id": 4280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996701.381, "dur": 4.012,
    "args": {
      "External id": 4281,"Record function id": 0, "Ev Idx": 3768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996702.612, "dur": 2.078,
    "args": {
      "External id": 4282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996703.229, "dur": 1.154,
    "args": {
      "External id": 4283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996703.619, "dur": 0.633,
    "args": {
      "External id": 4284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996708.428, "dur": 3.826,
    "args": {
      "External id": 4285,"Record function id": 0, "Ev Idx": 3772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996709.742, "dur": 1.820,
    "args": {
      "External id": 4286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996710.357, "dur": 0.874,
    "args": {
      "External id": 4287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996710.722, "dur": 0.347,
    "args": {
      "External id": 4288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996715.580, "dur": 52.861,
    "args": {
      "External id": 4289,"Record function id": 0, "Ev Idx": 3776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996716.844, "dur": 50.275,
    "args": {
      "External id": 4290,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996717.587, "dur": 48.638,
    "args": {
      "External id": 4291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996717.949, "dur": 45.305,
    "args": {
      "External id": 4292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996777.529, "dur": 8.159,
    "args": {
      "External id": 4293,"Record function id": 0, "Ev Idx": 3780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996782.631, "dur": 2.295,
    "args": {
      "External id": 4294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996783.615, "dur": 0.880,
    "args": {
      "External id": 4295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996784.031, "dur": 0.325,
    "args": {
      "External id": 4296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996788.850, "dur": 3.758,
    "args": {
      "External id": 4297,"Record function id": 0, "Ev Idx": 3784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996790.126, "dur": 1.775,
    "args": {
      "External id": 4298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996790.700, "dur": 0.881,
    "args": {
      "External id": 4299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996791.096, "dur": 0.335,
    "args": {
      "External id": 4300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996795.618, "dur": 3.859,
    "args": {
      "External id": 4301,"Record function id": 0, "Ev Idx": 3788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996796.976, "dur": 1.784,
    "args": {
      "External id": 4302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996797.635, "dur": 0.813,
    "args": {
      "External id": 4303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996797.994, "dur": 0.321,
    "args": {
      "External id": 4304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996802.504, "dur": 4.007,
    "args": {
      "External id": 4305,"Record function id": 0, "Ev Idx": 3792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996803.846, "dur": 1.953,
    "args": {
      "External id": 4306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996804.565, "dur": 0.927,
    "args": {
      "External id": 4307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996804.989, "dur": 0.345,
    "args": {
      "External id": 4308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996809.531, "dur": 3.764,
    "args": {
      "External id": 4309,"Record function id": 0, "Ev Idx": 3796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996810.824, "dur": 1.758,
    "args": {
      "External id": 4310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996811.440, "dur": 0.832,
    "args": {
      "External id": 4311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996811.820, "dur": 0.321,
    "args": {
      "External id": 4312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996817.008, "dur": 3.851,
    "args": {
      "External id": 4313,"Record function id": 0, "Ev Idx": 3800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996818.341, "dur": 1.820,
    "args": {
      "External id": 4314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996818.955, "dur": 0.881,
    "args": {
      "External id": 4315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996819.354, "dur": 0.353,
    "args": {
      "External id": 4316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996823.945, "dur": 3.691,
    "args": {
      "External id": 4317,"Record function id": 0, "Ev Idx": 3804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996825.174, "dur": 1.786,
    "args": {
      "External id": 4318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996825.807, "dur": 0.842,
    "args": {
      "External id": 4319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996826.187, "dur": 0.332,
    "args": {
      "External id": 4320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996830.650, "dur": 3.780,
    "args": {
      "External id": 4321,"Record function id": 0, "Ev Idx": 3808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996831.916, "dur": 1.815,
    "args": {
      "External id": 4322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996832.498, "dur": 0.923,
    "args": {
      "External id": 4323,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996832.867, "dur": 0.419,
    "args": {
      "External id": 4324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996837.457, "dur": 3.733,
    "args": {
      "External id": 4325,"Record function id": 0, "Ev Idx": 3812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996838.605, "dur": 1.910,
    "args": {
      "External id": 4326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996839.330, "dur": 0.873,
    "args": {
      "External id": 4327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996839.711, "dur": 0.357,
    "args": {
      "External id": 4328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996844.231, "dur": 3.905,
    "args": {
      "External id": 4329,"Record function id": 0, "Ev Idx": 3816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996845.456, "dur": 1.992,
    "args": {
      "External id": 4330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996846.301, "dur": 0.839,
    "args": {
      "External id": 4331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996846.685, "dur": 0.331,
    "args": {
      "External id": 4332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 3819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996851.118, "dur": 3.789,
    "args": {
      "External id": 4333,"Record function id": 0, "Ev Idx": 3820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996852.344, "dur": 1.871,
    "args": {
      "External id": 4334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996852.915, "dur": 0.990,
    "args": {
      "External id": 4335,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996853.419, "dur": 0.358,
    "args": {
      "External id": 4336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996858.036, "dur": 4.420,
    "args": {
      "External id": 4337,"Record function id": 0, "Ev Idx": 3824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996859.246, "dur": 2.504,
    "args": {
      "External id": 4338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996859.841, "dur": 1.560,
    "args": {
      "External id": 4339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996860.904, "dur": 0.341,
    "args": {
      "External id": 4340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 3827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996865.451, "dur": 6.065,
    "args": {
      "External id": 4341,"Record function id": 0, "Ev Idx": 3828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996866.697, "dur": 4.113,
    "args": {
      "External id": 4342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996867.316, "dur": 3.149,
    "args": {
      "External id": 4343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996867.686, "dur": 2.582,
    "args": {
      "External id": 4344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996874.555, "dur": 3.914,
    "args": {
      "External id": 4345,"Record function id": 0, "Ev Idx": 3832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996875.909, "dur": 1.819,
    "args": {
      "External id": 4346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996876.525, "dur": 0.892,
    "args": {
      "External id": 4347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996876.897, "dur": 0.390,
    "args": {
      "External id": 4348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996881.514, "dur": 3.963,
    "args": {
      "External id": 4349,"Record function id": 0, "Ev Idx": 3836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996882.879, "dur": 1.895,
    "args": {
      "External id": 4350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996883.582, "dur": 0.875,
    "args": {
      "External id": 4351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996883.945, "dur": 0.385,
    "args": {
      "External id": 4352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996888.461, "dur": 3.885,
    "args": {
      "External id": 4353,"Record function id": 0, "Ev Idx": 3840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996889.862, "dur": 1.816,
    "args": {
      "External id": 4354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 3841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996890.538, "dur": 0.827,
    "args": {
      "External id": 4355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 3842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996890.900, "dur": 0.340,
    "args": {
      "External id": 4356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 3843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996895.364, "dur": 51.922,
    "args": {
      "External id": 4357,"Record function id": 0, "Ev Idx": 3844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996896.604, "dur": 45.906,
    "args": {
      "External id": 4358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996897.279, "dur": 44.362,
    "args": {
      "External id": 4359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996897.648, "dur": 42.818,
    "args": {
      "External id": 4360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996956.239, "dur": 8.035,
    "args": {
      "External id": 4361,"Record function id": 0, "Ev Idx": 3848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996961.167, "dur": 2.351,
    "args": {
      "External id": 4362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996962.111, "dur": 0.949,
    "args": {
      "External id": 4363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996962.554, "dur": 0.352,
    "args": {
      "External id": 4364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996967.356, "dur": 5.507,
    "args": {
      "External id": 4365,"Record function id": 0, "Ev Idx": 3852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996968.678, "dur": 3.458,
    "args": {
      "External id": 4366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996970.842, "dur": 0.939,
    "args": {
      "External id": 4367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996971.227, "dur": 0.425,
    "args": {
      "External id": 4368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996976.056, "dur": 3.868,
    "args": {
      "External id": 4369,"Record function id": 0, "Ev Idx": 3856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996977.342, "dur": 1.881,
    "args": {
      "External id": 4370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996977.962, "dur": 0.942,
    "args": {
      "External id": 4371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996978.358, "dur": 0.390,
    "args": {
      "External id": 4372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996983.049, "dur": 3.816,
    "args": {
      "External id": 4373,"Record function id": 0, "Ev Idx": 3860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996984.417, "dur": 1.756,
    "args": {
      "External id": 4374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996984.987, "dur": 0.878,
    "args": {
      "External id": 4375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996985.377, "dur": 0.345,
    "args": {
      "External id": 4376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996989.861, "dur": 3.865,
    "args": {
      "External id": 4377,"Record function id": 0, "Ev Idx": 3864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996991.185, "dur": 1.855,
    "args": {
      "External id": 4378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996991.822, "dur": 0.908,
    "args": {
      "External id": 4379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996992.204, "dur": 0.374,
    "args": {
      "External id": 4380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996996.689, "dur": 3.838,
    "args": {
      "External id": 4381,"Record function id": 0, "Ev Idx": 3868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507996997.891, "dur": 1.948,
    "args": {
      "External id": 4382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996998.509, "dur": 1.022,
    "args": {
      "External id": 4383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507996998.884, "dur": 0.510,
    "args": {
      "External id": 4384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507997003.833, "dur": 176.805,
    "args": {
      "External id": 4385,"Record function id": 0, "Sequence number": 4692, "Fwd thread id": 1, "Ev Idx": 3872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507997005.481, "dur": 132.632,
    "args": {
      "External id": 4386,"Sequence number": 4692, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 3873
    }
  },
  {
    "ph": "f", "id": 10, "pid": 14565, "tid": 14640, "ts": 1290507997005.481,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290507997066.732, "dur": 61.997,
    "args": {
      "External id": 4387,"Record function id": 0, "scope": "user_scope", "Ev Idx": 3874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290507997148.303, "dur": 29.809,
    "args": {
      "External id": 4388,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 3875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507997186.181, "dur": 5.663,
    "args": {
      "External id": 4389,"Record function id": 0, "Ev Idx": 3876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290507997187.996, "dur": 2.967,
    "args": {
      "External id": 4390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 3877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290507997189.032, "dur": 1.606,
    "args": {
      "External id": 4391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 3878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290507997189.693, "dur": 0.791,
    "args": {
      "External id": 4392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 3879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507999699.879, "dur": 3778.894,
    "args": {
      "External id": 4393,"Record function id": 0, "Sequence number": 4695, "Fwd thread id": 1, "Ev Idx": 3880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290507999702.770, "dur": 3566.942,
    "args": {
      "External id": 4394,"Sequence number": 4695, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 3881
    }
  },
  {
    "ph": "f", "id": 11, "pid": 14565, "tid": 14640, "ts": 1290507999702.770,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290508000038.048, "dur": 3191.727,
    "args": {
      "External id": 4395,"Record function id": 0, "scope": "user_scope", "Ev Idx": 3882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290508000620.268, "dur": 31.989,
    "args": {
      "External id": 4396,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 3883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003502.763, "dur": 17.733,
    "args": {
      "External id": 4397,"Record function id": 0, "Ev Idx": 3884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003511.501, "dur": 7.533,
    "args": {
      "External id": 4398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003514.537, "dur": 3.797,
    "args": {
      "External id": 4399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003516.239, "dur": 1.912,
    "args": {
      "External id": 4400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003526.456, "dur": 4.867,
    "args": {
      "External id": 4401,"Record function id": 0, "Ev Idx": 3888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003528.105, "dur": 2.485,
    "args": {
      "External id": 4402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003528.986, "dur": 1.273,
    "args": {
      "External id": 4403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003529.628, "dur": 0.484,
    "args": {
      "External id": 4404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003538.304, "dur": 5.643,
    "args": {
      "External id": 4405,"Record function id": 0, "Ev Idx": 3892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003539.929, "dur": 3.135,
    "args": {
      "External id": 4406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003540.638, "dur": 2.116,
    "args": {
      "External id": 4407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003541.143, "dur": 1.479,
    "args": {
      "External id": 4408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003548.898, "dur": 4.903,
    "args": {
      "External id": 4409,"Record function id": 0, "Ev Idx": 3896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003550.241, "dur": 2.572,
    "args": {
      "External id": 4410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003550.946, "dur": 1.443,
    "args": {
      "External id": 4411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003551.591, "dur": 0.655,
    "args": {
      "External id": 4412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003558.737, "dur": 5.574,
    "args": {
      "External id": 4413,"Record function id": 0, "Ev Idx": 3900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003560.327, "dur": 3.134,
    "args": {
      "External id": 4414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003560.947, "dur": 1.953,
    "args": {
      "External id": 4415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003561.302, "dur": 1.298,
    "args": {
      "External id": 4416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003568.547, "dur": 5.062,
    "args": {
      "External id": 4417,"Record function id": 0, "Ev Idx": 3904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003570.103, "dur": 2.658,
    "args": {
      "External id": 4418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003570.712, "dur": 1.627,
    "args": {
      "External id": 4419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003571.188, "dur": 0.923,
    "args": {
      "External id": 4420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003581.088, "dur": 5.322,
    "args": {
      "External id": 4421,"Record function id": 0, "Ev Idx": 3908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003582.668, "dur": 2.925,
    "args": {
      "External id": 4422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003583.322, "dur": 1.943,
    "args": {
      "External id": 4423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003583.910, "dur": 0.943,
    "args": {
      "External id": 4424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003591.569, "dur": 5.794,
    "args": {
      "External id": 4425,"Record function id": 0, "Ev Idx": 3912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003593.146, "dur": 3.368,
    "args": {
      "External id": 4426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003593.784, "dur": 2.408,
    "args": {
      "External id": 4427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003594.182, "dur": 1.622,
    "args": {
      "External id": 4428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003603.721, "dur": 4.800,
    "args": {
      "External id": 4429,"Record function id": 0, "Ev Idx": 3916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003605.206, "dur": 2.408,
    "args": {
      "External id": 4430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003605.788, "dur": 1.193,
    "args": {
      "External id": 4431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003606.191, "dur": 0.519,
    "args": {
      "External id": 4432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003612.705, "dur": 5.823,
    "args": {
      "External id": 4433,"Record function id": 0, "Ev Idx": 3920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003614.130, "dur": 3.521,
    "args": {
      "External id": 4434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003615.008, "dur": 1.519,
    "args": {
      "External id": 4435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003615.394, "dur": 0.741,
    "args": {
      "External id": 4436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003624.961, "dur": 4.840,
    "args": {
      "External id": 4437,"Record function id": 0, "Ev Idx": 3924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003626.196, "dur": 2.778,
    "args": {
      "External id": 4438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003626.901, "dur": 1.363,
    "args": {
      "External id": 4439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003627.266, "dur": 0.593,
    "args": {
      "External id": 4440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003635.226, "dur": 6.462,
    "args": {
      "External id": 4441,"Record function id": 0, "Ev Idx": 3928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003636.863, "dur": 3.985,
    "args": {
      "External id": 4442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003637.499, "dur": 2.882,
    "args": {
      "External id": 4443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003637.870, "dur": 2.242,
    "args": {
      "External id": 4444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003648.780, "dur": 7.022,
    "args": {
      "External id": 4445,"Record function id": 0, "Ev Idx": 3932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003650.832, "dur": 4.130,
    "args": {
      "External id": 4446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003651.601, "dur": 2.714,
    "args": {
      "External id": 4447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003652.892, "dur": 1.033,
    "args": {
      "External id": 4448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003661.683, "dur": 4.797,
    "args": {
      "External id": 4449,"Record function id": 0, "Ev Idx": 3936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003663.128, "dur": 2.530,
    "args": {
      "External id": 4450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003663.807, "dur": 1.420,
    "args": {
      "External id": 4451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003664.189, "dur": 0.909,
    "args": {
      "External id": 4452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003673.267, "dur": 4.775,
    "args": {
      "External id": 4453,"Record function id": 0, "Ev Idx": 3940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003674.635, "dur": 2.547,
    "args": {
      "External id": 4454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003675.240, "dur": 1.499,
    "args": {
      "External id": 4455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003675.598, "dur": 0.859,
    "args": {
      "External id": 4456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003682.275, "dur": 5.043,
    "args": {
      "External id": 4457,"Record function id": 0, "Ev Idx": 3944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003683.762, "dur": 2.719,
    "args": {
      "External id": 4458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003684.480, "dur": 1.518,
    "args": {
      "External id": 4459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003684.868, "dur": 0.744,
    "args": {
      "External id": 4460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003692.620, "dur": 5.102,
    "args": {
      "External id": 4461,"Record function id": 0, "Ev Idx": 3948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003693.965, "dur": 2.915,
    "args": {
      "External id": 4462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003694.686, "dur": 1.610,
    "args": {
      "External id": 4463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003695.270, "dur": 0.627,
    "args": {
      "External id": 4464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003701.769, "dur": 5.549,
    "args": {
      "External id": 4465,"Record function id": 0, "Ev Idx": 3952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003703.533, "dur": 2.926,
    "args": {
      "External id": 4466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003704.126, "dur": 2.016,
    "args": {
      "External id": 4467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003704.699, "dur": 1.173,
    "args": {
      "External id": 4468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003712.679, "dur": 6.438,
    "args": {
      "External id": 4469,"Record function id": 0, "Ev Idx": 3956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003714.580, "dur": 3.686,
    "args": {
      "External id": 4470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003715.301, "dur": 2.326,
    "args": {
      "External id": 4471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003715.866, "dur": 1.350,
    "args": {
      "External id": 4472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003723.051, "dur": 151.665,
    "args": {
      "External id": 4473,"Record function id": 0, "Ev Idx": 3960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003724.791, "dur": 149.024,
    "args": {
      "External id": 4474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003726.375, "dur": 146.912,
    "args": {
      "External id": 4475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003871.601, "dur": 1.377,
    "args": {
      "External id": 4476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 3963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003882.055, "dur": 7.168,
    "args": {
      "External id": 4477,"Record function id": 0, "Ev Idx": 3964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003885.165, "dur": 3.181,
    "args": {
      "External id": 4478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003886.032, "dur": 1.558,
    "args": {
      "External id": 4479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003886.606, "dur": 0.592,
    "args": {
      "External id": 4480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003893.630, "dur": 7.139,
    "args": {
      "External id": 4481,"Record function id": 0, "Ev Idx": 3968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003896.812, "dur": 3.077,
    "args": {
      "External id": 4482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003897.557, "dur": 1.624,
    "args": {
      "External id": 4483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003897.934, "dur": 0.955,
    "args": {
      "External id": 4484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003908.287, "dur": 6.018,
    "args": {
      "External id": 4485,"Record function id": 0, "Ev Idx": 3972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003911.110, "dur": 2.350,
    "args": {
      "External id": 4486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003911.662, "dur": 1.491,
    "args": {
      "External id": 4487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003912.061, "dur": 0.699,
    "args": {
      "External id": 4488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 3975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003918.558, "dur": 6.531,
    "args": {
      "External id": 4489,"Record function id": 0, "Ev Idx": 3976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003921.544, "dur": 2.546,
    "args": {
      "External id": 4490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003922.153, "dur": 1.276,
    "args": {
      "External id": 4491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003922.518, "dur": 0.632,
    "args": {
      "External id": 4492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 3979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003930.378, "dur": 6.490,
    "args": {
      "External id": 4493,"Record function id": 0, "Ev Idx": 3980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003933.020, "dur": 3.013,
    "args": {
      "External id": 4494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003933.624, "dur": 2.102,
    "args": {
      "External id": 4495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003934.294, "dur": 0.874,
    "args": {
      "External id": 4496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 3983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003941.635, "dur": 5.188,
    "args": {
      "External id": 4497,"Record function id": 0, "Ev Idx": 3984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003943.496, "dur": 2.496,
    "args": {
      "External id": 4498,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003944.172, "dur": 1.270,
    "args": {
      "External id": 4499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003944.742, "dur": 0.450,
    "args": {
      "External id": 4500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 3987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003954.283, "dur": 9.265,
    "args": {
      "External id": 4501,"Record function id": 0, "Ev Idx": 3988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003957.999, "dur": 4.730,
    "args": {
      "External id": 4502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003958.668, "dur": 3.338,
    "args": {
      "External id": 4503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003959.493, "dur": 2.245,
    "args": {
      "External id": 4504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003968.677, "dur": 6.927,
    "args": {
      "External id": 4505,"Record function id": 0, "Ev Idx": 3992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003971.169, "dur": 3.601,
    "args": {
      "External id": 4506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003971.740, "dur": 2.149,
    "args": {
      "External id": 4507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003972.120, "dur": 1.239,
    "args": {
      "External id": 4508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 3995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003982.383, "dur": 6.759,
    "args": {
      "External id": 4509,"Record function id": 0, "Ev Idx": 3996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003985.629, "dur": 2.662,
    "args": {
      "External id": 4510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003986.216, "dur": 1.675,
    "args": {
      "External id": 4511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003986.581, "dur": 0.923,
    "args": {
      "External id": 4512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 3999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003993.428, "dur": 5.943,
    "args": {
      "External id": 4513,"Record function id": 0, "Ev Idx": 4000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508003996.312, "dur": 2.228,
    "args": {
      "External id": 4514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003996.910, "dur": 1.192,
    "args": {
      "External id": 4515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508003997.235, "dur": 0.602,
    "args": {
      "External id": 4516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004005.607, "dur": 6.994,
    "args": {
      "External id": 4517,"Record function id": 0, "Ev Idx": 4004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004008.413, "dur": 3.208,
    "args": {
      "External id": 4518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004009.012, "dur": 2.185,
    "args": {
      "External id": 4519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004009.380, "dur": 1.363,
    "args": {
      "External id": 4520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004018.039, "dur": 6.296,
    "args": {
      "External id": 4521,"Record function id": 0, "Ev Idx": 4008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004021.085, "dur": 2.416,
    "args": {
      "External id": 4522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004021.744, "dur": 1.309,
    "args": {
      "External id": 4523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004022.111, "dur": 0.610,
    "args": {
      "External id": 4524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004030.715, "dur": 6.084,
    "args": {
      "External id": 4525,"Record function id": 0, "Ev Idx": 4012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004033.203, "dur": 2.744,
    "args": {
      "External id": 4526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004034.027, "dur": 1.064,
    "args": {
      "External id": 4527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004034.390, "dur": 0.435,
    "args": {
      "External id": 4528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004041.418, "dur": 6.367,
    "args": {
      "External id": 4529,"Record function id": 0, "Ev Idx": 4016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004044.412, "dur": 2.515,
    "args": {
      "External id": 4530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004044.979, "dur": 1.473,
    "args": {
      "External id": 4531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004045.351, "dur": 0.815,
    "args": {
      "External id": 4532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004053.788, "dur": 5.029,
    "args": {
      "External id": 4533,"Record function id": 0, "Ev Idx": 4020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004056.397, "dur": 1.731,
    "args": {
      "External id": 4534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004057.018, "dur": 0.802,
    "args": {
      "External id": 4535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004057.380, "dur": 0.316,
    "args": {
      "External id": 4536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004062.496, "dur": 6.922,
    "args": {
      "External id": 4537,"Record function id": 0, "Ev Idx": 4024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004065.508, "dur": 3.075,
    "args": {
      "External id": 4538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004066.042, "dur": 2.090,
    "args": {
      "External id": 4539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004066.406, "dur": 1.326,
    "args": {
      "External id": 4540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004076.770, "dur": 9.367,
    "args": {
      "External id": 4541,"Record function id": 0, "Ev Idx": 4028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004083.170, "dur": 1.807,
    "args": {
      "External id": 4542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004083.841, "dur": 0.825,
    "args": {
      "External id": 4543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004084.214, "dur": 0.316,
    "args": {
      "External id": 4544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004091.559, "dur": 6.437,
    "args": {
      "External id": 4545,"Record function id": 0, "Ev Idx": 4032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004094.231, "dur": 2.764,
    "args": {
      "External id": 4546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004094.791, "dur": 1.777,
    "args": {
      "External id": 4547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004095.166, "dur": 0.917,
    "args": {
      "External id": 4548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004105.093, "dur": 6.537,
    "args": {
      "External id": 4549,"Record function id": 0, "Ev Idx": 4036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004107.209, "dur": 3.573,
    "args": {
      "External id": 4550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004107.850, "dur": 2.461,
    "args": {
      "External id": 4551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004108.332, "dur": 1.541,
    "args": {
      "External id": 4552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004116.239, "dur": 6.322,
    "args": {
      "External id": 4553,"Record function id": 0, "Ev Idx": 4040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004118.848, "dur": 2.853,
    "args": {
      "External id": 4554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004119.546, "dur": 1.710,
    "args": {
      "External id": 4555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004120.021, "dur": 0.955,
    "args": {
      "External id": 4556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004129.423, "dur": 5.742,
    "args": {
      "External id": 4557,"Record function id": 0, "Ev Idx": 4044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004131.722, "dur": 2.447,
    "args": {
      "External id": 4558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004132.449, "dur": 1.413,
    "args": {
      "External id": 4559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004132.835, "dur": 0.749,
    "args": {
      "External id": 4560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004139.632, "dur": 378.687,
    "args": {
      "External id": 4561,"Record function id": 0, "Ev Idx": 4048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004141.621, "dur": 375.599,
    "args": {
      "External id": 4562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004142.244, "dur": 374.337,
    "args": {
      "External id": 4563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004514.730, "dur": 1.435,
    "args": {
      "External id": 4564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004526.288, "dur": 6.525,
    "args": {
      "External id": 4565,"Record function id": 0, "Ev Idx": 4052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004528.971, "dur": 2.976,
    "args": {
      "External id": 4566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004529.724, "dur": 1.338,
    "args": {
      "External id": 4567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004530.092, "dur": 0.684,
    "args": {
      "External id": 4568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004537.951, "dur": 12.656,
    "args": {
      "External id": 4569,"Record function id": 0, "Ev Idx": 4056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004546.587, "dur": 3.177,
    "args": {
      "External id": 4570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004547.669, "dur": 1.525,
    "args": {
      "External id": 4571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004548.234, "dur": 0.693,
    "args": {
      "External id": 4572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004557.229, "dur": 6.643,
    "args": {
      "External id": 4573,"Record function id": 0, "Ev Idx": 4060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004559.914, "dur": 3.103,
    "args": {
      "External id": 4574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004560.597, "dur": 2.033,
    "args": {
      "External id": 4575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004561.203, "dur": 0.941,
    "args": {
      "External id": 4576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004568.530, "dur": 7.349,
    "args": {
      "External id": 4577,"Record function id": 0, "Ev Idx": 4064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004571.605, "dur": 3.408,
    "args": {
      "External id": 4578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004572.470, "dur": 2.095,
    "args": {
      "External id": 4579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004572.830, "dur": 1.469,
    "args": {
      "External id": 4580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004581.554, "dur": 7.194,
    "args": {
      "External id": 4581,"Record function id": 0, "Ev Idx": 4068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004584.349, "dur": 3.294,
    "args": {
      "External id": 4582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004585.100, "dur": 1.857,
    "args": {
      "External id": 4583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004585.575, "dur": 1.251,
    "args": {
      "External id": 4584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004593.123, "dur": 4.803,
    "args": {
      "External id": 4585,"Record function id": 0, "Ev Idx": 4072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004594.663, "dur": 2.412,
    "args": {
      "External id": 4586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004595.385, "dur": 1.135,
    "args": {
      "External id": 4587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004595.747, "dur": 0.646,
    "args": {
      "External id": 4588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004604.310, "dur": 6.657,
    "args": {
      "External id": 4589,"Record function id": 0, "Ev Idx": 4076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004607.372, "dur": 2.719,
    "args": {
      "External id": 4590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004608.002, "dur": 1.649,
    "args": {
      "External id": 4591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004608.388, "dur": 0.806,
    "args": {
      "External id": 4592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004616.538, "dur": 7.088,
    "args": {
      "External id": 4593,"Record function id": 0, "Ev Idx": 4080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004619.342, "dur": 3.210,
    "args": {
      "External id": 4594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004619.946, "dur": 2.213,
    "args": {
      "External id": 4595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004620.350, "dur": 1.680,
    "args": {
      "External id": 4596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004629.980, "dur": 7.161,
    "args": {
      "External id": 4597,"Record function id": 0, "Ev Idx": 4084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004633.208, "dur": 3.092,
    "args": {
      "External id": 4598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004633.877, "dur": 1.784,
    "args": {
      "External id": 4599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004634.413, "dur": 0.724,
    "args": {
      "External id": 4600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004641.855, "dur": 6.848,
    "args": {
      "External id": 4601,"Record function id": 0, "Ev Idx": 4088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004645.119, "dur": 2.731,
    "args": {
      "External id": 4602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004645.712, "dur": 1.690,
    "args": {
      "External id": 4603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004646.243, "dur": 0.932,
    "args": {
      "External id": 4604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004655.070, "dur": 6.926,
    "args": {
      "External id": 4605,"Record function id": 0, "Ev Idx": 4092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508004658.001, "dur": 3.137,
    "args": {
      "External id": 4606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004658.593, "dur": 2.100,
    "args": {
      "External id": 4607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508004659.115, "dur": 1.450,
    "args": {
      "External id": 4608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005184.353, "dur": 78.218,
    "args": {
      "External id": 4609,"Record function id": 0, "Ev Idx": 4096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005255.050, "dur": 5.848,
    "args": {
      "External id": 4610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005256.987, "dur": 3.147,
    "args": {
      "External id": 4611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005257.431, "dur": 2.410,
    "args": {
      "External id": 4612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005269.468, "dur": 6.801,
    "args": {
      "External id": 4613,"Record function id": 0, "Ev Idx": 4100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005271.805, "dur": 3.573,
    "args": {
      "External id": 4614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005272.668, "dur": 2.132,
    "args": {
      "External id": 4615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005273.161, "dur": 1.375,
    "args": {
      "External id": 4616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005280.182, "dur": 6.045,
    "args": {
      "External id": 4617,"Record function id": 0, "Ev Idx": 4104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005282.428, "dur": 2.984,
    "args": {
      "External id": 4618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005283.000, "dur": 1.988,
    "args": {
      "External id": 4619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005283.365, "dur": 1.352,
    "args": {
      "External id": 4620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005291.438, "dur": 5.468,
    "args": {
      "External id": 4621,"Record function id": 0, "Ev Idx": 4108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005292.949, "dur": 3.143,
    "args": {
      "External id": 4622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005293.556, "dur": 2.110,
    "args": {
      "External id": 4623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005293.922, "dur": 1.474,
    "args": {
      "External id": 4624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005300.541, "dur": 6.716,
    "args": {
      "External id": 4625,"Record function id": 0, "Ev Idx": 4112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005302.414, "dur": 4.011,
    "args": {
      "External id": 4626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005302.980, "dur": 3.010,
    "args": {
      "External id": 4627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005303.402, "dur": 2.316,
    "args": {
      "External id": 4628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005311.720, "dur": 6.727,
    "args": {
      "External id": 4629,"Record function id": 0, "Ev Idx": 4116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005313.693, "dur": 3.908,
    "args": {
      "External id": 4630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005314.290, "dur": 2.798,
    "args": {
      "External id": 4631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005315.499, "dur": 1.316,
    "args": {
      "External id": 4632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005322.948, "dur": 5.992,
    "args": {
      "External id": 4633,"Record function id": 0, "Ev Idx": 4120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005325.305, "dur": 2.825,
    "args": {
      "External id": 4634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005325.939, "dur": 1.756,
    "args": {
      "External id": 4635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005326.317, "dur": 0.784,
    "args": {
      "External id": 4636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005333.630, "dur": 6.025,
    "args": {
      "External id": 4637,"Record function id": 0, "Ev Idx": 4124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005335.341, "dur": 3.382,
    "args": {
      "External id": 4638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005335.905, "dur": 2.239,
    "args": {
      "External id": 4639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005336.275, "dur": 1.640,
    "args": {
      "External id": 4640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005343.683, "dur": 6.253,
    "args": {
      "External id": 4641,"Record function id": 0, "Ev Idx": 4128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005345.761, "dur": 3.372,
    "args": {
      "External id": 4642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005346.336, "dur": 2.146,
    "args": {
      "External id": 4643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005346.759, "dur": 1.153,
    "args": {
      "External id": 4644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005355.221, "dur": 5.597,
    "args": {
      "External id": 4645,"Record function id": 0, "Ev Idx": 4132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005356.894, "dur": 3.120,
    "args": {
      "External id": 4646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005357.486, "dur": 1.958,
    "args": {
      "External id": 4647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005357.885, "dur": 1.299,
    "args": {
      "External id": 4648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005365.278, "dur": 6.414,
    "args": {
      "External id": 4649,"Record function id": 0, "Ev Idx": 4136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005367.337, "dur": 3.505,
    "args": {
      "External id": 4650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005367.899, "dur": 2.483,
    "args": {
      "External id": 4651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005368.290, "dur": 1.567,
    "args": {
      "External id": 4652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005376.841, "dur": 6.046,
    "args": {
      "External id": 4653,"Record function id": 0, "Ev Idx": 4140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005378.883, "dur": 3.133,
    "args": {
      "External id": 4654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005379.551, "dur": 1.917,
    "args": {
      "External id": 4655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005379.906, "dur": 1.292,
    "args": {
      "External id": 4656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005387.375, "dur": 9.050,
    "args": {
      "External id": 4657,"Record function id": 0, "Ev Idx": 4144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005389.236, "dur": 6.345,
    "args": {
      "External id": 4658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005391.333, "dur": 3.564,
    "args": {
      "External id": 4659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005391.695, "dur": 2.922,
    "args": {
      "External id": 4660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005400.426, "dur": 6.244,
    "args": {
      "External id": 4661,"Record function id": 0, "Ev Idx": 4148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005402.599, "dur": 3.230,
    "args": {
      "External id": 4662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005403.186, "dur": 2.117,
    "args": {
      "External id": 4663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005403.601, "dur": 1.408,
    "args": {
      "External id": 4664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005410.416, "dur": 6.146,
    "args": {
      "External id": 4665,"Record function id": 0, "Ev Idx": 4152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005412.357, "dur": 3.404,
    "args": {
      "External id": 4666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005413.008, "dur": 2.252,
    "args": {
      "External id": 4667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005413.371, "dur": 1.605,
    "args": {
      "External id": 4668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005421.180, "dur": 6.189,
    "args": {
      "External id": 4669,"Record function id": 0, "Ev Idx": 4156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005423.317, "dur": 3.227,
    "args": {
      "External id": 4670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005423.909, "dur": 2.201,
    "args": {
      "External id": 4671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005424.302, "dur": 1.539,
    "args": {
      "External id": 4672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005431.326, "dur": 5.984,
    "args": {
      "External id": 4673,"Record function id": 0, "Ev Idx": 4160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005433.334, "dur": 3.170,
    "args": {
      "External id": 4674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005434.045, "dur": 2.035,
    "args": {
      "External id": 4675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005434.454, "dur": 1.365,
    "args": {
      "External id": 4676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005440.995, "dur": 5.775,
    "args": {
      "External id": 4677,"Record function id": 0, "Ev Idx": 4164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005443.183, "dur": 2.789,
    "args": {
      "External id": 4678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005443.767, "dur": 1.756,
    "args": {
      "External id": 4679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005444.134, "dur": 1.117,
    "args": {
      "External id": 4680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005451.489, "dur": 6.187,
    "args": {
      "External id": 4681,"Record function id": 0, "Ev Idx": 4168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005453.644, "dur": 3.212,
    "args": {
      "External id": 4682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005454.197, "dur": 2.036,
    "args": {
      "External id": 4683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005454.566, "dur": 1.435,
    "args": {
      "External id": 4684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005461.880, "dur": 6.848,
    "args": {
      "External id": 4685,"Record function id": 0, "Ev Idx": 4172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005464.126, "dur": 3.770,
    "args": {
      "External id": 4686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005464.863, "dur": 2.464,
    "args": {
      "External id": 4687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005465.274, "dur": 1.782,
    "args": {
      "External id": 4688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005472.932, "dur": 6.310,
    "args": {
      "External id": 4689,"Record function id": 0, "Ev Idx": 4176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005474.914, "dur": 3.520,
    "args": {
      "External id": 4690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005475.625, "dur": 2.256,
    "args": {
      "External id": 4691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005475.971, "dur": 1.635,
    "args": {
      "External id": 4692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005483.157, "dur": 5.466,
    "args": {
      "External id": 4693,"Record function id": 0, "Ev Idx": 4180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005484.734, "dur": 3.087,
    "args": {
      "External id": 4694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005485.284, "dur": 2.086,
    "args": {
      "External id": 4695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005485.653, "dur": 1.443,
    "args": {
      "External id": 4696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005492.403, "dur": 6.198,
    "args": {
      "External id": 4697,"Record function id": 0, "Ev Idx": 4184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005493.909, "dur": 3.857,
    "args": {
      "External id": 4698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005494.636, "dur": 2.276,
    "args": {
      "External id": 4699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005494.995, "dur": 1.648,
    "args": {
      "External id": 4700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005502.494, "dur": 5.385,
    "args": {
      "External id": 4701,"Record function id": 0, "Ev Idx": 4188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005504.091, "dur": 2.866,
    "args": {
      "External id": 4702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005504.699, "dur": 1.815,
    "args": {
      "External id": 4703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005505.224, "dur": 1.014,
    "args": {
      "External id": 4704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005512.382, "dur": 6.226,
    "args": {
      "External id": 4705,"Record function id": 0, "Ev Idx": 4192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005514.476, "dur": 3.232,
    "args": {
      "External id": 4706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005515.046, "dur": 2.347,
    "args": {
      "External id": 4707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005515.444, "dur": 1.702,
    "args": {
      "External id": 4708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005522.556, "dur": 6.387,
    "args": {
      "External id": 4709,"Record function id": 0, "Ev Idx": 4196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005525.010, "dur": 3.133,
    "args": {
      "External id": 4710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005525.591, "dur": 2.117,
    "args": {
      "External id": 4711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005525.966, "dur": 1.468,
    "args": {
      "External id": 4712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005533.749, "dur": 5.543,
    "args": {
      "External id": 4713,"Record function id": 0, "Ev Idx": 4200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005535.514, "dur": 2.975,
    "args": {
      "External id": 4714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005536.122, "dur": 1.947,
    "args": {
      "External id": 4715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005536.476, "dur": 1.327,
    "args": {
      "External id": 4716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005544.265, "dur": 5.773,
    "args": {
      "External id": 4717,"Record function id": 0, "Ev Idx": 4204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005545.933, "dur": 3.302,
    "args": {
      "External id": 4718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005546.536, "dur": 2.246,
    "args": {
      "External id": 4719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005546.946, "dur": 1.564,
    "args": {
      "External id": 4720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005554.933, "dur": 5.972,
    "args": {
      "External id": 4721,"Record function id": 0, "Ev Idx": 4208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005556.880, "dur": 3.127,
    "args": {
      "External id": 4722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005557.467, "dur": 2.087,
    "args": {
      "External id": 4723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005557.843, "dur": 1.446,
    "args": {
      "External id": 4724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005565.417, "dur": 5.734,
    "args": {
      "External id": 4725,"Record function id": 0, "Ev Idx": 4212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005567.319, "dur": 3.034,
    "args": {
      "External id": 4726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005567.895, "dur": 1.992,
    "args": {
      "External id": 4727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005568.250, "dur": 1.365,
    "args": {
      "External id": 4728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005575.995, "dur": 5.502,
    "args": {
      "External id": 4729,"Record function id": 0, "Ev Idx": 4216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005577.450, "dur": 3.259,
    "args": {
      "External id": 4730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005578.061, "dur": 1.979,
    "args": {
      "External id": 4731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005578.499, "dur": 1.188,
    "args": {
      "External id": 4732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005586.172, "dur": 5.845,
    "args": {
      "External id": 4733,"Record function id": 0, "Ev Idx": 4220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005587.728, "dur": 3.488,
    "args": {
      "External id": 4734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005588.345, "dur": 2.417,
    "args": {
      "External id": 4735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005588.702, "dur": 1.441,
    "args": {
      "External id": 4736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005596.239, "dur": 6.635,
    "args": {
      "External id": 4737,"Record function id": 0, "Ev Idx": 4224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005598.458, "dur": 3.608,
    "args": {
      "External id": 4738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005599.142, "dur": 2.355,
    "args": {
      "External id": 4739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005599.538, "dur": 1.694,
    "args": {
      "External id": 4740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005607.258, "dur": 6.137,
    "args": {
      "External id": 4741,"Record function id": 0, "Ev Idx": 4228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005609.478, "dur": 3.112,
    "args": {
      "External id": 4742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005610.144, "dur": 1.893,
    "args": {
      "External id": 4743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005610.506, "dur": 1.262,
    "args": {
      "External id": 4744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005617.300, "dur": 6.068,
    "args": {
      "External id": 4745,"Record function id": 0, "Ev Idx": 4232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005619.404, "dur": 2.977,
    "args": {
      "External id": 4746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005619.959, "dur": 1.985,
    "args": {
      "External id": 4747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005620.339, "dur": 1.338,
    "args": {
      "External id": 4748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005627.570, "dur": 5.618,
    "args": {
      "External id": 4749,"Record function id": 0, "Ev Idx": 4236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005629.625, "dur": 2.756,
    "args": {
      "External id": 4750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005630.229, "dur": 1.694,
    "args": {
      "External id": 4751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005630.588, "dur": 1.063,
    "args": {
      "External id": 4752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005636.956, "dur": 5.723,
    "args": {
      "External id": 4753,"Record function id": 0, "Ev Idx": 4240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005638.921, "dur": 2.950,
    "args": {
      "External id": 4754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005639.490, "dur": 1.939,
    "args": {
      "External id": 4755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005639.902, "dur": 1.259,
    "args": {
      "External id": 4756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005647.076, "dur": 5.469,
    "args": {
      "External id": 4757,"Record function id": 0, "Ev Idx": 4244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005649.026, "dur": 2.526,
    "args": {
      "External id": 4758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005649.564, "dur": 1.527,
    "args": {
      "External id": 4759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005649.926, "dur": 0.902,
    "args": {
      "External id": 4760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005657.020, "dur": 5.644,
    "args": {
      "External id": 4761,"Record function id": 0, "Ev Idx": 4248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005658.636, "dur": 3.219,
    "args": {
      "External id": 4762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005659.206, "dur": 2.227,
    "args": {
      "External id": 4763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005659.632, "dur": 1.518,
    "args": {
      "External id": 4764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005666.777, "dur": 4.187,
    "args": {
      "External id": 4765,"Record function id": 0, "Ev Idx": 4252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005668.224, "dur": 2.055,
    "args": {
      "External id": 4766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005668.834, "dur": 0.978,
    "args": {
      "External id": 4767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005669.219, "dur": 0.455,
    "args": {
      "External id": 4768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005675.246, "dur": 6.862,
    "args": {
      "External id": 4769,"Record function id": 0, "Ev Idx": 4256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005677.803, "dur": 3.117,
    "args": {
      "External id": 4770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005678.358, "dur": 2.140,
    "args": {
      "External id": 4771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005678.713, "dur": 1.515,
    "args": {
      "External id": 4772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005685.918, "dur": 5.592,
    "args": {
      "External id": 4773,"Record function id": 0, "Ev Idx": 4260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005687.325, "dur": 3.229,
    "args": {
      "External id": 4774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005687.879, "dur": 2.233,
    "args": {
      "External id": 4775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005688.264, "dur": 1.576,
    "args": {
      "External id": 4776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005695.863, "dur": 6.398,
    "args": {
      "External id": 4777,"Record function id": 0, "Ev Idx": 4264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005698.248, "dur": 3.188,
    "args": {
      "External id": 4778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 4265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005698.982, "dur": 1.994,
    "args": {
      "External id": 4779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 4266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005699.342, "dur": 1.391,
    "args": {
      "External id": 4780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 4267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005705.873, "dur": 4.286,
    "args": {
      "External id": 4781,"Record function id": 0, "Ev Idx": 4268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005707.285, "dur": 2.200,
    "args": {
      "External id": 4782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005707.896, "dur": 1.262,
    "args": {
      "External id": 4783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005708.273, "dur": 0.736,
    "args": {
      "External id": 4784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005714.282, "dur": 5.608,
    "args": {
      "External id": 4785,"Record function id": 0, "Ev Idx": 4272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005715.873, "dur": 3.211,
    "args": {
      "External id": 4786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005716.505, "dur": 2.135,
    "args": {
      "External id": 4787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005716.897, "dur": 1.471,
    "args": {
      "External id": 4788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005723.224, "dur": 5.602,
    "args": {
      "External id": 4789,"Record function id": 0, "Ev Idx": 4276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005724.932, "dur": 3.075,
    "args": {
      "External id": 4790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005725.477, "dur": 2.103,
    "args": {
      "External id": 4791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005725.840, "dur": 1.472,
    "args": {
      "External id": 4792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005732.566, "dur": 5.744,
    "args": {
      "External id": 4793,"Record function id": 0, "Ev Idx": 4280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005734.577, "dur": 2.931,
    "args": {
      "External id": 4794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005735.114, "dur": 1.956,
    "args": {
      "External id": 4795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005735.486, "dur": 1.317,
    "args": {
      "External id": 4796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005742.424, "dur": 4.934,
    "args": {
      "External id": 4797,"Record function id": 0, "Ev Idx": 4284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005743.890, "dur": 2.656,
    "args": {
      "External id": 4798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005744.477, "dur": 1.615,
    "args": {
      "External id": 4799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005744.850, "dur": 0.973,
    "args": {
      "External id": 4800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005751.734, "dur": 6.395,
    "args": {
      "External id": 4801,"Record function id": 0, "Ev Idx": 4288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005754.267, "dur": 3.039,
    "args": {
      "External id": 4802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 4289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005754.830, "dur": 2.041,
    "args": {
      "External id": 4803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 4290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005755.207, "dur": 1.399,
    "args": {
      "External id": 4804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 4291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005763.513, "dur": 6.107,
    "args": {
      "External id": 4805,"Record function id": 0, "Ev Idx": 4292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005765.578, "dur": 3.242,
    "args": {
      "External id": 4806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005766.158, "dur": 2.232,
    "args": {
      "External id": 4807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005766.558, "dur": 1.566,
    "args": {
      "External id": 4808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005774.372, "dur": 6.420,
    "args": {
      "External id": 4809,"Record function id": 0, "Ev Idx": 4296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005776.293, "dur": 3.683,
    "args": {
      "External id": 4810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005776.946, "dur": 2.458,
    "args": {
      "External id": 4811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005777.425, "dur": 1.716,
    "args": {
      "External id": 4812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005784.849, "dur": 6.360,
    "args": {
      "External id": 4813,"Record function id": 0, "Ev Idx": 4300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005786.356, "dur": 4.037,
    "args": {
      "External id": 4814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005787.129, "dur": 2.688,
    "args": {
      "External id": 4815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005787.542, "dur": 1.828,
    "args": {
      "External id": 4816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005794.890, "dur": 6.165,
    "args": {
      "External id": 4817,"Record function id": 0, "Ev Idx": 4304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005797.005, "dur": 3.240,
    "args": {
      "External id": 4818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005797.535, "dur": 2.252,
    "args": {
      "External id": 4819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005797.899, "dur": 1.613,
    "args": {
      "External id": 4820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005804.704, "dur": 6.762,
    "args": {
      "External id": 4821,"Record function id": 0, "Ev Idx": 4308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005806.977, "dur": 3.409,
    "args": {
      "External id": 4822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005807.586, "dur": 2.369,
    "args": {
      "External id": 4823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005807.946, "dur": 1.743,
    "args": {
      "External id": 4824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005815.834, "dur": 5.323,
    "args": {
      "External id": 4825,"Record function id": 0, "Ev Idx": 4312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005817.604, "dur": 2.714,
    "args": {
      "External id": 4826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005818.188, "dur": 1.684,
    "args": {
      "External id": 4827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005818.570, "dur": 1.031,
    "args": {
      "External id": 4828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005825.701, "dur": 6.230,
    "args": {
      "External id": 4829,"Record function id": 0, "Ev Idx": 4316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005827.804, "dur": 3.325,
    "args": {
      "External id": 4830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005828.412, "dur": 2.050,
    "args": {
      "External id": 4831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005828.797, "dur": 1.394,
    "args": {
      "External id": 4832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005836.273, "dur": 5.037,
    "args": {
      "External id": 4833,"Record function id": 0, "Ev Idx": 4320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005837.906, "dur": 2.583,
    "args": {
      "External id": 4834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005838.513, "dur": 1.551,
    "args": {
      "External id": 4835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005838.884, "dur": 0.910,
    "args": {
      "External id": 4836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005846.258, "dur": 5.674,
    "args": {
      "External id": 4837,"Record function id": 0, "Ev Idx": 4324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005847.714, "dur": 3.247,
    "args": {
      "External id": 4838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005848.288, "dur": 1.976,
    "args": {
      "External id": 4839,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005848.660, "dur": 1.338,
    "args": {
      "External id": 4840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005856.087, "dur": 4.100,
    "args": {
      "External id": 4841,"Record function id": 0, "Ev Idx": 4328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005857.393, "dur": 2.093,
    "args": {
      "External id": 4842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005858.018, "dur": 1.144,
    "args": {
      "External id": 4843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005858.379, "dur": 0.653,
    "args": {
      "External id": 4844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005864.278, "dur": 5.895,
    "args": {
      "External id": 4845,"Record function id": 0, "Ev Idx": 4332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005865.582, "dur": 3.755,
    "args": {
      "External id": 4846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005866.186, "dur": 2.698,
    "args": {
      "External id": 4847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005866.553, "dur": 2.047,
    "args": {
      "External id": 4848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005874.708, "dur": 5.824,
    "args": {
      "External id": 4849,"Record function id": 0, "Ev Idx": 4336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005876.233, "dur": 3.504,
    "args": {
      "External id": 4850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005876.953, "dur": 2.332,
    "args": {
      "External id": 4851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005877.309, "dur": 1.732,
    "args": {
      "External id": 4852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005884.775, "dur": 4.518,
    "args": {
      "External id": 4853,"Record function id": 0, "Ev Idx": 4340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005886.064, "dur": 2.406,
    "args": {
      "External id": 4854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005886.750, "dur": 1.154,
    "args": {
      "External id": 4855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005887.129, "dur": 0.625,
    "args": {
      "External id": 4856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005893.080, "dur": 5.528,
    "args": {
      "External id": 4857,"Record function id": 0, "Ev Idx": 4344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005894.366, "dur": 3.457,
    "args": {
      "External id": 4858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005895.082, "dur": 2.191,
    "args": {
      "External id": 4859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005895.442, "dur": 1.251,
    "args": {
      "External id": 4860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005902.261, "dur": 4.645,
    "args": {
      "External id": 4861,"Record function id": 0, "Ev Idx": 4348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005903.553, "dur": 2.668,
    "args": {
      "External id": 4862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005904.125, "dur": 1.615,
    "args": {
      "External id": 4863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005904.503, "dur": 1.106,
    "args": {
      "External id": 4864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005910.240, "dur": 4.114,
    "args": {
      "External id": 4865,"Record function id": 0, "Ev Idx": 4352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005911.498, "dur": 2.180,
    "args": {
      "External id": 4866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005912.112, "dur": 1.246,
    "args": {
      "External id": 4867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005912.486, "dur": 0.746,
    "args": {
      "External id": 4868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005917.898, "dur": 4.395,
    "args": {
      "External id": 4869,"Record function id": 0, "Ev Idx": 4356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005919.291, "dur": 2.320,
    "args": {
      "External id": 4870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005919.899, "dur": 1.395,
    "args": {
      "External id": 4871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005920.281, "dur": 0.878,
    "args": {
      "External id": 4872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005925.843, "dur": 4.188,
    "args": {
      "External id": 4873,"Record function id": 0, "Ev Idx": 4360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005927.013, "dur": 2.322,
    "args": {
      "External id": 4874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005927.652, "dur": 1.361,
    "args": {
      "External id": 4875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005928.010, "dur": 0.879,
    "args": {
      "External id": 4876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005933.404, "dur": 3.775,
    "args": {
      "External id": 4877,"Record function id": 0, "Ev Idx": 4364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005934.581, "dur": 1.887,
    "args": {
      "External id": 4878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005935.180, "dur": 0.964,
    "args": {
      "External id": 4879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005935.549, "dur": 0.464,
    "args": {
      "External id": 4880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005940.324, "dur": 3.948,
    "args": {
      "External id": 4881,"Record function id": 0, "Ev Idx": 4368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005941.597, "dur": 1.980,
    "args": {
      "External id": 4882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005942.174, "dur": 1.088,
    "args": {
      "External id": 4883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005942.542, "dur": 0.592,
    "args": {
      "External id": 4884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005947.479, "dur": 4.038,
    "args": {
      "External id": 4885,"Record function id": 0, "Ev Idx": 4372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005948.815, "dur": 1.884,
    "args": {
      "External id": 4886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005949.402, "dur": 0.974,
    "args": {
      "External id": 4887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005949.799, "dur": 0.429,
    "args": {
      "External id": 4888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005954.721, "dur": 3.712,
    "args": {
      "External id": 4889,"Record function id": 0, "Ev Idx": 4376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005955.923, "dur": 1.807,
    "args": {
      "External id": 4890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005956.526, "dur": 0.887,
    "args": {
      "External id": 4891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005956.927, "dur": 0.341,
    "args": {
      "External id": 4892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005961.721, "dur": 3.793,
    "args": {
      "External id": 4893,"Record function id": 0, "Ev Idx": 4380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005962.946, "dur": 1.850,
    "args": {
      "External id": 4894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005963.626, "dur": 0.851,
    "args": {
      "External id": 4895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005963.985, "dur": 0.350,
    "args": {
      "External id": 4896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005969.006, "dur": 3.968,
    "args": {
      "External id": 4897,"Record function id": 0, "Ev Idx": 4384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005970.281, "dur": 1.982,
    "args": {
      "External id": 4898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005971.009, "dur": 0.823,
    "args": {
      "External id": 4899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005971.368, "dur": 0.322,
    "args": {
      "External id": 4900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005976.372, "dur": 4.494,
    "args": {
      "External id": 4901,"Record function id": 0, "Ev Idx": 4388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005977.589, "dur": 2.575,
    "args": {
      "External id": 4902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005978.280, "dur": 1.454,
    "args": {
      "External id": 4903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005978.672, "dur": 0.915,
    "args": {
      "External id": 4904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005984.383, "dur": 3.857,
    "args": {
      "External id": 4905,"Record function id": 0, "Ev Idx": 4392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005985.533, "dur": 2.013,
    "args": {
      "External id": 4906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005986.138, "dur": 1.089,
    "args": {
      "External id": 4907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508005986.510, "dur": 0.569,
    "args": {
      "External id": 4908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005991.442, "dur": 162.905,
    "args": {
      "External id": 4909,"Record function id": 0, "Ev Idx": 4396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508005992.712, "dur": 160.893,
    "args": {
      "External id": 4910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006151.890, "dur": 1.347,
    "args": {
      "External id": 4911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006152.390, "dur": 0.703,
    "args": {
      "External id": 4912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006157.793, "dur": 3.639,
    "args": {
      "External id": 4913,"Record function id": 0, "Ev Idx": 4400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006159.015, "dur": 1.725,
    "args": {
      "External id": 4914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006159.599, "dur": 0.815,
    "args": {
      "External id": 4915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006159.982, "dur": 0.300,
    "args": {
      "External id": 4916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006164.649, "dur": 3.892,
    "args": {
      "External id": 4917,"Record function id": 0, "Ev Idx": 4404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006166.028, "dur": 1.816,
    "args": {
      "External id": 4918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006166.649, "dur": 0.870,
    "args": {
      "External id": 4919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006167.023, "dur": 0.347,
    "args": {
      "External id": 4920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006171.735, "dur": 3.821,
    "args": {
      "External id": 4921,"Record function id": 0, "Ev Idx": 4408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006173.120, "dur": 1.736,
    "args": {
      "External id": 4922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006173.673, "dur": 0.853,
    "args": {
      "External id": 4923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006174.078, "dur": 0.317,
    "args": {
      "External id": 4924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006178.791, "dur": 3.702,
    "args": {
      "External id": 4925,"Record function id": 0, "Ev Idx": 4412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006180.084, "dur": 1.735,
    "args": {
      "External id": 4926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006180.672, "dur": 0.823,
    "args": {
      "External id": 4927,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006181.047, "dur": 0.309,
    "args": {
      "External id": 4928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006185.651, "dur": 3.790,
    "args": {
      "External id": 4929,"Record function id": 0, "Ev Idx": 4416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006187.091, "dur": 1.670,
    "args": {
      "External id": 4930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006187.661, "dur": 0.779,
    "args": {
      "External id": 4931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006188.015, "dur": 0.297,
    "args": {
      "External id": 4932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006192.581, "dur": 3.661,
    "args": {
      "External id": 4933,"Record function id": 0, "Ev Idx": 4420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006193.832, "dur": 1.734,
    "args": {
      "External id": 4934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006194.471, "dur": 0.771,
    "args": {
      "External id": 4935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006194.840, "dur": 0.273,
    "args": {
      "External id": 4936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006199.533, "dur": 4.699,
    "args": {
      "External id": 4937,"Record function id": 0, "Ev Idx": 4424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006201.911, "dur": 1.642,
    "args": {
      "External id": 4938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006202.451, "dur": 0.784,
    "args": {
      "External id": 4939,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006202.812, "dur": 0.287,
    "args": {
      "External id": 4940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006207.458, "dur": 3.977,
    "args": {
      "External id": 4941,"Record function id": 0, "Ev Idx": 4428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006208.840, "dur": 1.915,
    "args": {
      "External id": 4942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006209.472, "dur": 0.879,
    "args": {
      "External id": 4943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006209.840, "dur": 0.377,
    "args": {
      "External id": 4944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006214.720, "dur": 4.179,
    "args": {
      "External id": 4945,"Record function id": 0, "Ev Idx": 4432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006216.204, "dur": 1.995,
    "args": {
      "External id": 4946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006216.907, "dur": 0.822,
    "args": {
      "External id": 4947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006217.319, "dur": 0.277,
    "args": {
      "External id": 4948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006222.377, "dur": 3.871,
    "args": {
      "External id": 4949,"Record function id": 0, "Ev Idx": 4436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006223.810, "dur": 1.758,
    "args": {
      "External id": 4950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006224.503, "dur": 0.743,
    "args": {
      "External id": 4951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006224.854, "dur": 0.265,
    "args": {
      "External id": 4952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006229.403, "dur": 3.642,
    "args": {
      "External id": 4953,"Record function id": 0, "Ev Idx": 4440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006230.625, "dur": 1.747,
    "args": {
      "External id": 4954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006231.200, "dur": 0.841,
    "args": {
      "External id": 4955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006231.600, "dur": 0.315,
    "args": {
      "External id": 4956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006236.584, "dur": 3.679,
    "args": {
      "External id": 4957,"Record function id": 0, "Ev Idx": 4444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006237.886, "dur": 1.696,
    "args": {
      "External id": 4958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006238.500, "dur": 0.760,
    "args": {
      "External id": 4959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006238.852, "dur": 0.278,
    "args": {
      "External id": 4960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006243.491, "dur": 3.826,
    "args": {
      "External id": 4961,"Record function id": 0, "Ev Idx": 4448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006244.787, "dur": 1.856,
    "args": {
      "External id": 4962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006245.357, "dur": 0.938,
    "args": {
      "External id": 4963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006245.724, "dur": 0.437,
    "args": {
      "External id": 4964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006250.813, "dur": 3.894,
    "args": {
      "External id": 4965,"Record function id": 0, "Ev Idx": 4452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006252.182, "dur": 1.847,
    "args": {
      "External id": 4966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006252.744, "dur": 0.962,
    "args": {
      "External id": 4967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006253.115, "dur": 0.470,
    "args": {
      "External id": 4968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006257.987, "dur": 3.745,
    "args": {
      "External id": 4969,"Record function id": 0, "Ev Idx": 4456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006259.213, "dur": 1.829,
    "args": {
      "External id": 4970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006259.768, "dur": 0.952,
    "args": {
      "External id": 4971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006260.123, "dur": 0.461,
    "args": {
      "External id": 4972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006265.600, "dur": 4.040,
    "args": {
      "External id": 4973,"Record function id": 0, "Ev Idx": 4460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006267.006, "dur": 1.954,
    "args": {
      "External id": 4974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006267.635, "dur": 1.007,
    "args": {
      "External id": 4975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006268.047, "dur": 0.460,
    "args": {
      "External id": 4976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006273.099, "dur": 4.009,
    "args": {
      "External id": 4977,"Record function id": 0, "Ev Idx": 4464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006274.628, "dur": 1.806,
    "args": {
      "External id": 4978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006275.349, "dur": 0.763,
    "args": {
      "External id": 4979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006275.707, "dur": 0.274,
    "args": {
      "External id": 4980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006280.753, "dur": 3.947,
    "args": {
      "External id": 4981,"Record function id": 0, "Ev Idx": 4468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006282.099, "dur": 1.890,
    "args": {
      "External id": 4982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006282.812, "dur": 0.763,
    "args": {
      "External id": 4983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006283.173, "dur": 0.272,
    "args": {
      "External id": 4984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006288.236, "dur": 3.997,
    "args": {
      "External id": 4985,"Record function id": 0, "Ev Idx": 4472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006289.582, "dur": 1.962,
    "args": {
      "External id": 4986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006290.255, "dur": 0.851,
    "args": {
      "External id": 4987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006290.608, "dur": 0.367,
    "args": {
      "External id": 4988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006295.590, "dur": 3.919,
    "args": {
      "External id": 4989,"Record function id": 0, "Ev Idx": 4476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006297.079, "dur": 1.755,
    "args": {
      "External id": 4990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006297.668, "dur": 0.832,
    "args": {
      "External id": 4991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006298.074, "dur": 0.295,
    "args": {
      "External id": 4992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006302.857, "dur": 3.642,
    "args": {
      "External id": 4993,"Record function id": 0, "Ev Idx": 4480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006304.109, "dur": 1.707,
    "args": {
      "External id": 4994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006304.728, "dur": 0.768,
    "args": {
      "External id": 4995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006305.084, "dur": 0.285,
    "args": {
      "External id": 4996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006309.751, "dur": 3.762,
    "args": {
      "External id": 4997,"Record function id": 0, "Ev Idx": 4484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006311.181, "dur": 1.648,
    "args": {
      "External id": 4998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006311.728, "dur": 0.778,
    "args": {
      "External id": 4999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006312.089, "dur": 0.284,
    "args": {
      "External id": 5000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006316.661, "dur": 3.612,
    "args": {
      "External id": 5001,"Record function id": 0, "Ev Idx": 4488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006317.877, "dur": 1.720,
    "args": {
      "External id": 5002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006318.491, "dur": 0.778,
    "args": {
      "External id": 5003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006318.848, "dur": 0.290,
    "args": {
      "External id": 5004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006323.381, "dur": 3.622,
    "args": {
      "External id": 5005,"Record function id": 0, "Ev Idx": 4492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006324.585, "dur": 1.739,
    "args": {
      "External id": 5006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006325.149, "dur": 0.848,
    "args": {
      "External id": 5007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006325.513, "dur": 0.354,
    "args": {
      "External id": 5008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006330.227, "dur": 3.579,
    "args": {
      "External id": 5009,"Record function id": 0, "Ev Idx": 4496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006331.426, "dur": 1.680,
    "args": {
      "External id": 5010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006331.986, "dur": 0.796,
    "args": {
      "External id": 5011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006332.373, "dur": 0.278,
    "args": {
      "External id": 5012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006337.076, "dur": 3.767,
    "args": {
      "External id": 5013,"Record function id": 0, "Ev Idx": 4500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006338.496, "dur": 1.631,
    "args": {
      "External id": 5014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006339.045, "dur": 0.762,
    "args": {
      "External id": 5015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006339.393, "dur": 0.287,
    "args": {
      "External id": 5016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006344.148, "dur": 3.642,
    "args": {
      "External id": 5017,"Record function id": 0, "Ev Idx": 4504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006345.328, "dur": 1.759,
    "args": {
      "External id": 5018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006345.945, "dur": 0.828,
    "args": {
      "External id": 5019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006346.335, "dur": 0.307,
    "args": {
      "External id": 5020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006351.291, "dur": 4.021,
    "args": {
      "External id": 5021,"Record function id": 0, "Ev Idx": 4508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006352.924, "dur": 1.700,
    "args": {
      "External id": 5022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006353.488, "dur": 0.818,
    "args": {
      "External id": 5023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006353.893, "dur": 0.284,
    "args": {
      "External id": 5024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006358.970, "dur": 4.074,
    "args": {
      "External id": 5025,"Record function id": 0, "Ev Idx": 4512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006360.264, "dur": 2.066,
    "args": {
      "External id": 5026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006361.042, "dur": 0.830,
    "args": {
      "External id": 5027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006361.436, "dur": 0.293,
    "args": {
      "External id": 5028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006366.434, "dur": 3.997,
    "args": {
      "External id": 5029,"Record function id": 0, "Ev Idx": 4516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006367.791, "dur": 1.929,
    "args": {
      "External id": 5030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006368.512, "dur": 0.773,
    "args": {
      "External id": 5031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006368.874, "dur": 0.282,
    "args": {
      "External id": 5032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006373.565, "dur": 4.102,
    "args": {
      "External id": 5033,"Record function id": 0, "Ev Idx": 4520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006375.034, "dur": 1.936,
    "args": {
      "External id": 5034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006375.633, "dur": 1.013,
    "args": {
      "External id": 5035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006376.007, "dur": 0.507,
    "args": {
      "External id": 5036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006381.034, "dur": 4.092,
    "args": {
      "External id": 5037,"Record function id": 0, "Ev Idx": 4524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006382.489, "dur": 1.946,
    "args": {
      "External id": 5038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006383.079, "dur": 1.025,
    "args": {
      "External id": 5039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006383.436, "dur": 0.540,
    "args": {
      "External id": 5040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006388.346, "dur": 3.797,
    "args": {
      "External id": 5041,"Record function id": 0, "Ev Idx": 4528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006389.742, "dur": 1.709,
    "args": {
      "External id": 5042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006390.340, "dur": 0.792,
    "args": {
      "External id": 5043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006390.703, "dur": 0.299,
    "args": {
      "External id": 5044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006395.363, "dur": 3.779,
    "args": {
      "External id": 5045,"Record function id": 0, "Ev Idx": 4532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006396.730, "dur": 1.739,
    "args": {
      "External id": 5046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006397.380, "dur": 0.772,
    "args": {
      "External id": 5047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006397.739, "dur": 0.284,
    "args": {
      "External id": 5048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006402.304, "dur": 3.790,
    "args": {
      "External id": 5049,"Record function id": 0, "Ev Idx": 4536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006403.699, "dur": 1.688,
    "args": {
      "External id": 5050,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006404.291, "dur": 0.778,
    "args": {
      "External id": 5051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006404.664, "dur": 0.273,
    "args": {
      "External id": 5052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006409.345, "dur": 3.729,
    "args": {
      "External id": 5053,"Record function id": 0, "Ev Idx": 4540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006410.653, "dur": 1.746,
    "args": {
      "External id": 5054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006411.276, "dur": 0.805,
    "args": {
      "External id": 5055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006411.682, "dur": 0.269,
    "args": {
      "External id": 5056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006416.325, "dur": 3.673,
    "args": {
      "External id": 5057,"Record function id": 0, "Ev Idx": 4544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006417.595, "dur": 1.731,
    "args": {
      "External id": 5058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006418.217, "dur": 0.794,
    "args": {
      "External id": 5059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006418.593, "dur": 0.287,
    "args": {
      "External id": 5060,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006423.105, "dur": 3.905,
    "args": {
      "External id": 5061,"Record function id": 0, "Ev Idx": 4548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006424.565, "dur": 1.762,
    "args": {
      "External id": 5062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006425.188, "dur": 0.821,
    "args": {
      "External id": 5063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006425.544, "dur": 0.294,
    "args": {
      "External id": 5064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006430.440, "dur": 3.932,
    "args": {
      "External id": 5065,"Record function id": 0, "Ev Idx": 4552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006431.687, "dur": 1.975,
    "args": {
      "External id": 5066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006432.418, "dur": 0.797,
    "args": {
      "External id": 5067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006432.799, "dur": 0.289,
    "args": {
      "External id": 5068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006437.795, "dur": 4.153,
    "args": {
      "External id": 5069,"Record function id": 0, "Ev Idx": 4556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006439.170, "dur": 2.055,
    "args": {
      "External id": 5070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006439.931, "dur": 0.855,
    "args": {
      "External id": 5071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006440.304, "dur": 0.353,
    "args": {
      "External id": 5072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006445.337, "dur": 3.781,
    "args": {
      "External id": 5073,"Record function id": 0, "Ev Idx": 4560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006446.728, "dur": 1.710,
    "args": {
      "External id": 5074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006447.340, "dur": 0.778,
    "args": {
      "External id": 5075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006447.705, "dur": 0.287,
    "args": {
      "External id": 5076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006452.417, "dur": 3.865,
    "args": {
      "External id": 5077,"Record function id": 0, "Ev Idx": 4564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006453.845, "dur": 1.702,
    "args": {
      "External id": 5078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006454.445, "dur": 0.782,
    "args": {
      "External id": 5079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006454.820, "dur": 0.275,
    "args": {
      "External id": 5080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006459.521, "dur": 3.568,
    "args": {
      "External id": 5081,"Record function id": 0, "Ev Idx": 4568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006460.732, "dur": 1.685,
    "args": {
      "External id": 5082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006461.322, "dur": 0.771,
    "args": {
      "External id": 5083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006461.691, "dur": 0.271,
    "args": {
      "External id": 5084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006466.185, "dur": 3.562,
    "args": {
      "External id": 5085,"Record function id": 0, "Ev Idx": 4572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006467.351, "dur": 1.727,
    "args": {
      "External id": 5086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006467.985, "dur": 0.779,
    "args": {
      "External id": 5087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006468.352, "dur": 0.280,
    "args": {
      "External id": 5088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006473.290, "dur": 3.751,
    "args": {
      "External id": 5089,"Record function id": 0, "Ev Idx": 4576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006474.468, "dur": 1.853,
    "args": {
      "External id": 5090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006475.021, "dur": 0.982,
    "args": {
      "External id": 5091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006475.393, "dur": 0.475,
    "args": {
      "External id": 5092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006480.305, "dur": 3.538,
    "args": {
      "External id": 5093,"Record function id": 0, "Ev Idx": 4580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006481.504, "dur": 1.673,
    "args": {
      "External id": 5094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006482.064, "dur": 0.792,
    "args": {
      "External id": 5095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006482.440, "dur": 0.286,
    "args": {
      "External id": 5096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006487.319, "dur": 3.654,
    "args": {
      "External id": 5097,"Record function id": 0, "Ev Idx": 4584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006488.530, "dur": 1.767,
    "args": {
      "External id": 5098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006489.172, "dur": 0.804,
    "args": {
      "External id": 5099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006489.551, "dur": 0.293,
    "args": {
      "External id": 5100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006494.130, "dur": 3.681,
    "args": {
      "External id": 5101,"Record function id": 0, "Ev Idx": 4588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006495.311, "dur": 1.827,
    "args": {
      "External id": 5102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006495.923, "dur": 0.886,
    "args": {
      "External id": 5103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006496.291, "dur": 0.389,
    "args": {
      "External id": 5104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006501.064, "dur": 3.606,
    "args": {
      "External id": 5105,"Record function id": 0, "Ev Idx": 4592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006502.318, "dur": 1.698,
    "args": {
      "External id": 5106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 4593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006502.933, "dur": 0.762,
    "args": {
      "External id": 5107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 4594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006503.285, "dur": 0.281,
    "args": {
      "External id": 5108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 4595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006507.945, "dur": 3.625,
    "args": {
      "External id": 5109,"Record function id": 0, "Ev Idx": 4596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006509.188, "dur": 1.711,
    "args": {
      "External id": 5110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006509.757, "dur": 0.821,
    "args": {
      "External id": 5111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006510.158, "dur": 0.290,
    "args": {
      "External id": 5112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006515.369, "dur": 3.758,
    "args": {
      "External id": 5113,"Record function id": 0, "Ev Idx": 4600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006516.684, "dur": 1.772,
    "args": {
      "External id": 5114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006517.309, "dur": 0.826,
    "args": {
      "External id": 5115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006517.667, "dur": 0.307,
    "args": {
      "External id": 5116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 4603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006522.503, "dur": 3.725,
    "args": {
      "External id": 5117,"Record function id": 0, "Ev Idx": 4604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006523.784, "dur": 1.776,
    "args": {
      "External id": 5118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006524.426, "dur": 0.811,
    "args": {
      "External id": 5119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006524.783, "dur": 0.317,
    "args": {
      "External id": 5120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006731.967, "dur": 10.799,
    "args": {
      "External id": 5121,"Record function id": 0, "Ev Idx": 4608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006739.221, "dur": 2.639,
    "args": {
      "External id": 5122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006740.136, "dur": 1.127,
    "args": {
      "External id": 5123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006740.591, "dur": 0.507,
    "args": {
      "External id": 5124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006746.564, "dur": 4.036,
    "args": {
      "External id": 5125,"Record function id": 0, "Ev Idx": 4612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006748.056, "dur": 1.834,
    "args": {
      "External id": 5126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006748.729, "dur": 0.822,
    "args": {
      "External id": 5127,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006749.115, "dur": 0.299,
    "args": {
      "External id": 5128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006753.790, "dur": 4.037,
    "args": {
      "External id": 5129,"Record function id": 0, "Ev Idx": 4616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006755.262, "dur": 1.860,
    "args": {
      "External id": 5130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 4617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006755.957, "dur": 0.859,
    "args": {
      "External id": 5131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 4618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006756.333, "dur": 0.356,
    "args": {
      "External id": 5132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 4619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006761.026, "dur": 4.679,
    "args": {
      "External id": 5133,"Record function id": 0, "Ev Idx": 4620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006762.657, "dur": 2.292,
    "args": {
      "External id": 5134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006763.461, "dur": 0.912,
    "args": {
      "External id": 5135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006763.837, "dur": 0.395,
    "args": {
      "External id": 5136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006769.655, "dur": 4.128,
    "args": {
      "External id": 5137,"Record function id": 0, "Ev Idx": 4624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006770.919, "dur": 2.146,
    "args": {
      "External id": 5138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006771.639, "dur": 0.975,
    "args": {
      "External id": 5139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006772.012, "dur": 0.475,
    "args": {
      "External id": 5140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006777.211, "dur": 3.624,
    "args": {
      "External id": 5141,"Record function id": 0, "Ev Idx": 4628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006778.443, "dur": 1.686,
    "args": {
      "External id": 5142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006779.008, "dur": 0.818,
    "args": {
      "External id": 5143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006779.401, "dur": 0.299,
    "args": {
      "External id": 5144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006784.729, "dur": 3.740,
    "args": {
      "External id": 5145,"Record function id": 0, "Ev Idx": 4632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006785.960, "dur": 1.806,
    "args": {
      "External id": 5146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006786.537, "dur": 0.928,
    "args": {
      "External id": 5147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006786.923, "dur": 0.408,
    "args": {
      "External id": 5148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006791.763, "dur": 4.961,
    "args": {
      "External id": 5149,"Record function id": 0, "Ev Idx": 4636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006794.272, "dur": 1.731,
    "args": {
      "External id": 5150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006794.833, "dur": 0.860,
    "args": {
      "External id": 5151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006795.241, "dur": 0.308,
    "args": {
      "External id": 5152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006799.993, "dur": 3.770,
    "args": {
      "External id": 5153,"Record function id": 0, "Ev Idx": 4640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006801.362, "dur": 1.718,
    "args": {
      "External id": 5154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006801.933, "dur": 0.847,
    "args": {
      "External id": 5155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006802.318, "dur": 0.335,
    "args": {
      "External id": 5156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006807.107, "dur": 3.825,
    "args": {
      "External id": 5157,"Record function id": 0, "Ev Idx": 4644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006808.505, "dur": 1.725,
    "args": {
      "External id": 5158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006809.082, "dur": 0.843,
    "args": {
      "External id": 5159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006809.470, "dur": 0.319,
    "args": {
      "External id": 5160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508006814.407, "dur": 173.270,
    "args": {
      "External id": 5161,"Record function id": 0, "Sequence number": 4694, "Fwd thread id": 1, "Ev Idx": 4648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508006816.172, "dur": 133.133,
    "args": {
      "External id": 5162,"Sequence number": 4694, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 4649
    }
  },
  {
    "ph": "f", "id": 12, "pid": 14565, "tid": 14640, "ts": 1290508006816.172,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290508006879.543, "dur": 60.650,
    "args": {
      "External id": 5163,"Record function id": 0, "scope": "user_scope", "Ev Idx": 4650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290508006956.633, "dur": 28.341,
    "args": {
      "External id": 5164,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 4651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006993.019, "dur": 5.583,
    "args": {
      "External id": 5165,"Record function id": 0, "Ev Idx": 4652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508006994.798, "dur": 2.965,
    "args": {
      "External id": 5166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 4653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006995.769, "dur": 1.651,
    "args": {
      "External id": 5167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 4654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508006996.293, "dur": 0.980,
    "args": {
      "External id": 5168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 4655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508009500.264, "dur": 1560.479,
    "args": {
      "External id": 5169,"Record function id": 0, "Sequence number": 4697, "Fwd thread id": 1, "Ev Idx": 4656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508009502.814, "dur": 1471.419,
    "args": {
      "External id": 5170,"Sequence number": 4697, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 4657
    }
  },
  {
    "ph": "f", "id": 13, "pid": 14565, "tid": 14640, "ts": 1290508009502.814,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290508009622.799, "dur": 1330.864,
    "args": {
      "External id": 5171,"Record function id": 0, "scope": "user_scope", "Ev Idx": 4658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290508010178.765, "dur": 22.283,
    "args": {
      "External id": 5172,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 4659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011072.231, "dur": 14.457,
    "args": {
      "External id": 5173,"Record function id": 0, "Ev Idx": 4660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011075.009, "dur": 10.425,
    "args": {
      "External id": 5174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011077.452, "dur": 7.351,
    "args": {
      "External id": 5175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011078.362, "dur": 6.247,
    "args": {
      "External id": 5176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011090.866, "dur": 4.621,
    "args": {
      "External id": 5177,"Record function id": 0, "Ev Idx": 4664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011092.290, "dur": 2.404,
    "args": {
      "External id": 5178,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011093.100, "dur": 1.227,
    "args": {
      "External id": 5179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011093.797, "dur": 0.373,
    "args": {
      "External id": 5180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011099.124, "dur": 4.318,
    "args": {
      "External id": 5181,"Record function id": 0, "Ev Idx": 4668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011100.391, "dur": 2.257,
    "args": {
      "External id": 5182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011101.165, "dur": 1.147,
    "args": {
      "External id": 5183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011101.760, "dur": 0.343,
    "args": {
      "External id": 5184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011106.947, "dur": 3.954,
    "args": {
      "External id": 5185,"Record function id": 0, "Ev Idx": 4672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011108.242, "dur": 1.909,
    "args": {
      "External id": 5186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011108.846, "dur": 0.976,
    "args": {
      "External id": 5187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011109.298, "dur": 0.386,
    "args": {
      "External id": 5188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011114.234, "dur": 3.919,
    "args": {
      "External id": 5189,"Record function id": 0, "Ev Idx": 4676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011115.443, "dur": 1.985,
    "args": {
      "External id": 5190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011116.203, "dur": 0.877,
    "args": {
      "External id": 5191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011116.576, "dur": 0.347,
    "args": {
      "External id": 5192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011121.640, "dur": 4.050,
    "args": {
      "External id": 5193,"Record function id": 0, "Ev Idx": 4680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011122.936, "dur": 2.010,
    "args": {
      "External id": 5194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011123.630, "dur": 0.981,
    "args": {
      "External id": 5195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011124.022, "dur": 0.440,
    "args": {
      "External id": 5196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011128.952, "dur": 3.931,
    "args": {
      "External id": 5197,"Record function id": 0, "Ev Idx": 4684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011130.265, "dur": 1.894,
    "args": {
      "External id": 5198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011130.989, "dur": 0.845,
    "args": {
      "External id": 5199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011131.327, "dur": 0.367,
    "args": {
      "External id": 5200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011136.221, "dur": 4.242,
    "args": {
      "External id": 5201,"Record function id": 0, "Ev Idx": 4688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011137.439, "dur": 2.279,
    "args": {
      "External id": 5202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011138.058, "dur": 1.296,
    "args": {
      "External id": 5203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011138.421, "dur": 0.792,
    "args": {
      "External id": 5204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011143.805, "dur": 3.922,
    "args": {
      "External id": 5205,"Record function id": 0, "Ev Idx": 4692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011145.028, "dur": 1.977,
    "args": {
      "External id": 5206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011145.615, "dur": 1.061,
    "args": {
      "External id": 5207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011145.950, "dur": 0.581,
    "args": {
      "External id": 5208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011151.079, "dur": 4.802,
    "args": {
      "External id": 5209,"Record function id": 0, "Ev Idx": 4696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011152.350, "dur": 2.722,
    "args": {
      "External id": 5210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011153.182, "dur": 1.070,
    "args": {
      "External id": 5211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011153.559, "dur": 0.555,
    "args": {
      "External id": 5212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011159.165, "dur": 4.321,
    "args": {
      "External id": 5213,"Record function id": 0, "Ev Idx": 4700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011160.465, "dur": 2.232,
    "args": {
      "External id": 5214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011161.218, "dur": 1.046,
    "args": {
      "External id": 5215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011161.573, "dur": 0.548,
    "args": {
      "External id": 5216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011166.785, "dur": 4.228,
    "args": {
      "External id": 5217,"Record function id": 0, "Ev Idx": 4704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011168.111, "dur": 2.169,
    "args": {
      "External id": 5218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011168.764, "dur": 1.185,
    "args": {
      "External id": 5219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011169.118, "dur": 0.705,
    "args": {
      "External id": 5220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011174.275, "dur": 3.885,
    "args": {
      "External id": 5221,"Record function id": 0, "Ev Idx": 4708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011175.531, "dur": 1.886,
    "args": {
      "External id": 5222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011176.172, "dur": 0.932,
    "args": {
      "External id": 5223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011176.505, "dur": 0.460,
    "args": {
      "External id": 5224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011181.437, "dur": 3.909,
    "args": {
      "External id": 5225,"Record function id": 0, "Ev Idx": 4712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011182.635, "dur": 1.999,
    "args": {
      "External id": 5226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011183.265, "dur": 1.034,
    "args": {
      "External id": 5227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011183.674, "dur": 0.481,
    "args": {
      "External id": 5228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011188.717, "dur": 4.046,
    "args": {
      "External id": 5229,"Record function id": 0, "Ev Idx": 4716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011189.977, "dur": 2.062,
    "args": {
      "External id": 5230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011190.602, "dur": 1.097,
    "args": {
      "External id": 5231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011190.939, "dur": 0.617,
    "args": {
      "External id": 5232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011196.006, "dur": 4.050,
    "args": {
      "External id": 5233,"Record function id": 0, "Ev Idx": 4720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011197.380, "dur": 1.957,
    "args": {
      "External id": 5234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011197.995, "dur": 1.013,
    "args": {
      "External id": 5235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011198.361, "dur": 0.509,
    "args": {
      "External id": 5236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011203.305, "dur": 3.983,
    "args": {
      "External id": 5237,"Record function id": 0, "Ev Idx": 4724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011204.649, "dur": 1.933,
    "args": {
      "External id": 5238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011205.311, "dur": 0.945,
    "args": {
      "External id": 5239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011205.712, "dur": 0.398,
    "args": {
      "External id": 5240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011210.619, "dur": 4.574,
    "args": {
      "External id": 5241,"Record function id": 0, "Ev Idx": 4728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011212.016, "dur": 2.450,
    "args": {
      "External id": 5242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011212.601, "dur": 1.535,
    "args": {
      "External id": 5243,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011213.152, "dur": 0.842,
    "args": {
      "External id": 5244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011218.855, "dur": 4.251,
    "args": {
      "External id": 5245,"Record function id": 0, "Ev Idx": 4732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011220.295, "dur": 2.096,
    "args": {
      "External id": 5246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011220.893, "dur": 1.160,
    "args": {
      "External id": 5247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011221.316, "dur": 0.595,
    "args": {
      "External id": 5248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011226.347, "dur": 4.355,
    "args": {
      "External id": 5249,"Record function id": 0, "Ev Idx": 4736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011227.527, "dur": 2.457,
    "args": {
      "External id": 5250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011228.105, "dur": 1.547,
    "args": {
      "External id": 5251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011228.703, "dur": 0.807,
    "args": {
      "External id": 5252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011234.042, "dur": 4.786,
    "args": {
      "External id": 5253,"Record function id": 0, "Ev Idx": 4740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011235.464, "dur": 2.590,
    "args": {
      "External id": 5254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011236.433, "dur": 1.179,
    "args": {
      "External id": 5255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011236.877, "dur": 0.600,
    "args": {
      "External id": 5256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011242.351, "dur": 5.137,
    "args": {
      "External id": 5257,"Record function id": 0, "Ev Idx": 4744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011244.144, "dur": 2.610,
    "args": {
      "External id": 5258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011245.109, "dur": 1.225,
    "args": {
      "External id": 5259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011245.661, "dur": 0.531,
    "args": {
      "External id": 5260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011250.978, "dur": 3.794,
    "args": {
      "External id": 5261,"Record function id": 0, "Ev Idx": 4748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011252.226, "dur": 1.838,
    "args": {
      "External id": 5262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011252.826, "dur": 0.905,
    "args": {
      "External id": 5263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011253.225, "dur": 0.361,
    "args": {
      "External id": 5264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011258.062, "dur": 7.915,
    "args": {
      "External id": 5265,"Record function id": 0, "Ev Idx": 4752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011259.458, "dur": 5.780,
    "args": {
      "External id": 5266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011260.090, "dur": 4.837,
    "args": {
      "External id": 5267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011260.444, "dur": 4.344,
    "args": {
      "External id": 5268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011269.419, "dur": 3.897,
    "args": {
      "External id": 5269,"Record function id": 0, "Ev Idx": 4756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011270.654, "dur": 1.953,
    "args": {
      "External id": 5270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011271.274, "dur": 0.999,
    "args": {
      "External id": 5271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011271.749, "dur": 0.380,
    "args": {
      "External id": 5272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011276.639, "dur": 3.863,
    "args": {
      "External id": 5273,"Record function id": 0, "Ev Idx": 4760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011277.890, "dur": 1.899,
    "args": {
      "External id": 5274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011278.518, "dur": 0.941,
    "args": {
      "External id": 5275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011278.863, "dur": 0.452,
    "args": {
      "External id": 5276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011283.753, "dur": 3.887,
    "args": {
      "External id": 5277,"Record function id": 0, "Ev Idx": 4764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011285.005, "dur": 1.909,
    "args": {
      "External id": 5278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011285.599, "dur": 0.985,
    "args": {
      "External id": 5279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011286.071, "dur": 0.374,
    "args": {
      "External id": 5280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011291.006, "dur": 4.125,
    "args": {
      "External id": 5281,"Record function id": 0, "Ev Idx": 4768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011292.359, "dur": 2.058,
    "args": {
      "External id": 5282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011292.951, "dur": 1.138,
    "args": {
      "External id": 5283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011293.477, "dur": 0.464,
    "args": {
      "External id": 5284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011298.397, "dur": 3.926,
    "args": {
      "External id": 5285,"Record function id": 0, "Ev Idx": 4772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011299.567, "dur": 2.043,
    "args": {
      "External id": 5286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011300.167, "dur": 1.100,
    "args": {
      "External id": 5287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011300.516, "dur": 0.611,
    "args": {
      "External id": 5288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011305.603, "dur": 4.028,
    "args": {
      "External id": 5289,"Record function id": 0, "Ev Idx": 4776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011306.890, "dur": 2.011,
    "args": {
      "External id": 5290,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011307.553, "dur": 1.013,
    "args": {
      "External id": 5291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011307.956, "dur": 0.462,
    "args": {
      "External id": 5292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011312.944, "dur": 3.981,
    "args": {
      "External id": 5293,"Record function id": 0, "Ev Idx": 4780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011314.174, "dur": 2.041,
    "args": {
      "External id": 5294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011314.790, "dur": 1.097,
    "args": {
      "External id": 5295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011315.197, "dur": 0.547,
    "args": {
      "External id": 5296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011320.302, "dur": 4.141,
    "args": {
      "External id": 5297,"Record function id": 0, "Ev Idx": 4784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011321.504, "dur": 2.204,
    "args": {
      "External id": 5298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011322.217, "dur": 1.052,
    "args": {
      "External id": 5299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011322.580, "dur": 0.548,
    "args": {
      "External id": 5300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011327.726, "dur": 4.779,
    "args": {
      "External id": 5301,"Record function id": 0, "Ev Idx": 4788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011329.005, "dur": 2.757,
    "args": {
      "External id": 5302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011329.754, "dur": 1.577,
    "args": {
      "External id": 5303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011330.150, "dur": 1.040,
    "args": {
      "External id": 5304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011335.875, "dur": 4.922,
    "args": {
      "External id": 5305,"Record function id": 0, "Ev Idx": 4792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011337.173, "dur": 2.902,
    "args": {
      "External id": 5306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011337.841, "dur": 1.893,
    "args": {
      "External id": 5307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011338.390, "dur": 1.203,
    "args": {
      "External id": 5308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011344.304, "dur": 4.196,
    "args": {
      "External id": 5309,"Record function id": 0, "Ev Idx": 4796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011345.587, "dur": 2.200,
    "args": {
      "External id": 5310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011346.156, "dur": 1.292,
    "args": {
      "External id": 5311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011346.549, "dur": 0.753,
    "args": {
      "External id": 5312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011351.747, "dur": 4.208,
    "args": {
      "External id": 5313,"Record function id": 0, "Ev Idx": 4800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011352.954, "dur": 2.274,
    "args": {
      "External id": 5314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011353.573, "dur": 1.326,
    "args": {
      "External id": 5315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011353.981, "dur": 0.772,
    "args": {
      "External id": 5316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011359.650, "dur": 4.431,
    "args": {
      "External id": 5317,"Record function id": 0, "Ev Idx": 4804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011360.975, "dur": 2.378,
    "args": {
      "External id": 5318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011361.563, "dur": 1.457,
    "args": {
      "External id": 5319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011361.915, "dur": 0.968,
    "args": {
      "External id": 5320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011367.305, "dur": 3.926,
    "args": {
      "External id": 5321,"Record function id": 0, "Ev Idx": 4808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011368.513, "dur": 2.011,
    "args": {
      "External id": 5322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011369.265, "dur": 0.934,
    "args": {
      "External id": 5323,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011369.620, "dur": 0.435,
    "args": {
      "External id": 5324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011374.643, "dur": 4.165,
    "args": {
      "External id": 5325,"Record function id": 0, "Ev Idx": 4812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011375.873, "dur": 2.208,
    "args": {
      "External id": 5326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011376.538, "dur": 1.211,
    "args": {
      "External id": 5327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011376.885, "dur": 0.724,
    "args": {
      "External id": 5328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011382.202, "dur": 3.868,
    "args": {
      "External id": 5329,"Record function id": 0, "Ev Idx": 4816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011383.429, "dur": 1.916,
    "args": {
      "External id": 5330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011384.055, "dur": 0.938,
    "args": {
      "External id": 5331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011384.395, "dur": 0.457,
    "args": {
      "External id": 5332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011389.307, "dur": 3.874,
    "args": {
      "External id": 5333,"Record function id": 0, "Ev Idx": 4820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011390.528, "dur": 1.945,
    "args": {
      "External id": 5334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011391.133, "dur": 1.007,
    "args": {
      "External id": 5335,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011391.528, "dur": 0.467,
    "args": {
      "External id": 5336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011396.427, "dur": 3.921,
    "args": {
      "External id": 5337,"Record function id": 0, "Ev Idx": 4824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011397.620, "dur": 1.980,
    "args": {
      "External id": 5338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011398.236, "dur": 1.033,
    "args": {
      "External id": 5339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011398.592, "dur": 0.536,
    "args": {
      "External id": 5340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011403.702, "dur": 4.089,
    "args": {
      "External id": 5341,"Record function id": 0, "Ev Idx": 4828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011404.851, "dur": 2.206,
    "args": {
      "External id": 5342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011405.606, "dur": 0.954,
    "args": {
      "External id": 5343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011405.963, "dur": 0.456,
    "args": {
      "External id": 5344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011411.105, "dur": 4.027,
    "args": {
      "External id": 5345,"Record function id": 0, "Ev Idx": 4832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011412.291, "dur": 2.103,
    "args": {
      "External id": 5346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011413.010, "dur": 0.961,
    "args": {
      "External id": 5347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011413.377, "dur": 0.406,
    "args": {
      "External id": 5348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011418.380, "dur": 3.863,
    "args": {
      "External id": 5349,"Record function id": 0, "Ev Idx": 4836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011419.674, "dur": 1.841,
    "args": {
      "External id": 5350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011420.268, "dur": 0.924,
    "args": {
      "External id": 5351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011420.666, "dur": 0.381,
    "args": {
      "External id": 5352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011425.507, "dur": 5.915,
    "args": {
      "External id": 5353,"Record function id": 0, "Ev Idx": 4840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011428.826, "dur": 1.881,
    "args": {
      "External id": 5354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011429.447, "dur": 0.922,
    "args": {
      "External id": 5355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011429.844, "dur": 0.383,
    "args": {
      "External id": 5356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011434.721, "dur": 3.725,
    "args": {
      "External id": 5357,"Record function id": 0, "Ev Idx": 4844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011435.927, "dur": 1.811,
    "args": {
      "External id": 5358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011436.524, "dur": 0.877,
    "args": {
      "External id": 5359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011436.903, "dur": 0.351,
    "args": {
      "External id": 5360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011441.823, "dur": 3.994,
    "args": {
      "External id": 5361,"Record function id": 0, "Ev Idx": 4848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011443.135, "dur": 1.945,
    "args": {
      "External id": 5362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011443.758, "dur": 1.009,
    "args": {
      "External id": 5363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011444.131, "dur": 0.496,
    "args": {
      "External id": 5364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011449.242, "dur": 3.775,
    "args": {
      "External id": 5365,"Record function id": 0, "Ev Idx": 4852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011450.469, "dur": 1.838,
    "args": {
      "External id": 5366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011451.072, "dur": 0.906,
    "args": {
      "External id": 5367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011451.447, "dur": 0.387,
    "args": {
      "External id": 5368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011456.363, "dur": 3.787,
    "args": {
      "External id": 5369,"Record function id": 0, "Ev Idx": 4856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011457.555, "dur": 1.859,
    "args": {
      "External id": 5370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011458.173, "dur": 0.927,
    "args": {
      "External id": 5371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011458.647, "dur": 0.314,
    "args": {
      "External id": 5372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011463.486, "dur": 4.017,
    "args": {
      "External id": 5373,"Record function id": 0, "Ev Idx": 4860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011464.710, "dur": 2.081,
    "args": {
      "External id": 5374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011465.292, "dur": 1.169,
    "args": {
      "External id": 5375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011465.670, "dur": 0.646,
    "args": {
      "External id": 5376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011470.800, "dur": 3.992,
    "args": {
      "External id": 5377,"Record function id": 0, "Ev Idx": 4864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011472.035, "dur": 2.032,
    "args": {
      "External id": 5378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011472.665, "dur": 1.070,
    "args": {
      "External id": 5379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011473.083, "dur": 0.509,
    "args": {
      "External id": 5380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011478.042, "dur": 4.260,
    "args": {
      "External id": 5381,"Record function id": 0, "Ev Idx": 4868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011479.295, "dur": 2.289,
    "args": {
      "External id": 5382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011479.910, "dur": 1.347,
    "args": {
      "External id": 5383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011480.276, "dur": 0.837,
    "args": {
      "External id": 5384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011485.580, "dur": 4.130,
    "args": {
      "External id": 5385,"Record function id": 0, "Ev Idx": 4872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011486.817, "dur": 2.144,
    "args": {
      "External id": 5386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011487.518, "dur": 1.001,
    "args": {
      "External id": 5387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011487.884, "dur": 0.494,
    "args": {
      "External id": 5388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011493.078, "dur": 4.294,
    "args": {
      "External id": 5389,"Record function id": 0, "Ev Idx": 4876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011494.296, "dur": 2.334,
    "args": {
      "External id": 5390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011495.111, "dur": 1.090,
    "args": {
      "External id": 5391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011495.542, "dur": 0.519,
    "args": {
      "External id": 5392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011500.690, "dur": 3.990,
    "args": {
      "External id": 5393,"Record function id": 0, "Ev Idx": 4880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011501.892, "dur": 2.055,
    "args": {
      "External id": 5394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011502.500, "dur": 1.119,
    "args": {
      "External id": 5395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011503.007, "dur": 0.475,
    "args": {
      "External id": 5396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011507.931, "dur": 3.833,
    "args": {
      "External id": 5397,"Record function id": 0, "Ev Idx": 4884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011509.134, "dur": 1.880,
    "args": {
      "External id": 5398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011509.731, "dur": 0.956,
    "args": {
      "External id": 5399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011510.119, "dur": 0.429,
    "args": {
      "External id": 5400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011514.986, "dur": 3.857,
    "args": {
      "External id": 5401,"Record function id": 0, "Ev Idx": 4888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011516.146, "dur": 1.971,
    "args": {
      "External id": 5402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011516.877, "dur": 0.908,
    "args": {
      "External id": 5403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011517.274, "dur": 0.377,
    "args": {
      "External id": 5404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011522.049, "dur": 3.721,
    "args": {
      "External id": 5405,"Record function id": 0, "Ev Idx": 4892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011523.237, "dur": 1.816,
    "args": {
      "External id": 5406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011523.805, "dur": 0.925,
    "args": {
      "External id": 5407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011524.178, "dur": 0.411,
    "args": {
      "External id": 5408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011529.003, "dur": 4.103,
    "args": {
      "External id": 5409,"Record function id": 0, "Ev Idx": 4896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011530.297, "dur": 2.075,
    "args": {
      "External id": 5410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011530.902, "dur": 1.129,
    "args": {
      "External id": 5411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011531.397, "dur": 0.491,
    "args": {
      "External id": 5412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011536.317, "dur": 4.196,
    "args": {
      "External id": 5413,"Record function id": 0, "Ev Idx": 4900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011537.589, "dur": 2.220,
    "args": {
      "External id": 5414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011538.269, "dur": 1.215,
    "args": {
      "External id": 5415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011538.749, "dur": 0.590,
    "args": {
      "External id": 5416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011543.713, "dur": 3.880,
    "args": {
      "External id": 5417,"Record function id": 0, "Ev Idx": 4904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011544.964, "dur": 1.899,
    "args": {
      "External id": 5418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011545.585, "dur": 0.947,
    "args": {
      "External id": 5419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011545.971, "dur": 0.415,
    "args": {
      "External id": 5420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011550.815, "dur": 3.831,
    "args": {
      "External id": 5421,"Record function id": 0, "Ev Idx": 4908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011552.012, "dur": 1.922,
    "args": {
      "External id": 5422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011552.618, "dur": 0.990,
    "args": {
      "External id": 5423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011552.999, "dur": 0.456,
    "args": {
      "External id": 5424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011557.891, "dur": 3.969,
    "args": {
      "External id": 5425,"Record function id": 0, "Ev Idx": 4912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011559.163, "dur": 1.973,
    "args": {
      "External id": 5426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011559.760, "dur": 1.049,
    "args": {
      "External id": 5427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011560.132, "dur": 0.530,
    "args": {
      "External id": 5428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011565.082, "dur": 4.399,
    "args": {
      "External id": 5429,"Record function id": 0, "Ev Idx": 4916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011566.280, "dur": 2.466,
    "args": {
      "External id": 5430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011567.022, "dur": 1.277,
    "args": {
      "External id": 5431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011567.387, "dur": 0.770,
    "args": {
      "External id": 5432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011572.690, "dur": 4.228,
    "args": {
      "External id": 5433,"Record function id": 0, "Ev Idx": 4920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011573.898, "dur": 2.278,
    "args": {
      "External id": 5434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011574.606, "dur": 1.141,
    "args": {
      "External id": 5435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011574.975, "dur": 0.633,
    "args": {
      "External id": 5436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011580.113, "dur": 3.705,
    "args": {
      "External id": 5437,"Record function id": 0, "Ev Idx": 4924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011581.278, "dur": 1.827,
    "args": {
      "External id": 5438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011581.890, "dur": 0.891,
    "args": {
      "External id": 5439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011582.265, "dur": 0.372,
    "args": {
      "External id": 5440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011586.961, "dur": 4.018,
    "args": {
      "External id": 5441,"Record function id": 0, "Ev Idx": 4928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011588.215, "dur": 2.036,
    "args": {
      "External id": 5442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011588.848, "dur": 1.072,
    "args": {
      "External id": 5443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011589.210, "dur": 0.570,
    "args": {
      "External id": 5444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011594.204, "dur": 4.073,
    "args": {
      "External id": 5445,"Record function id": 0, "Ev Idx": 4932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011595.355, "dur": 2.212,
    "args": {
      "External id": 5446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011595.931, "dur": 1.308,
    "args": {
      "External id": 5447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011596.314, "dur": 0.750,
    "args": {
      "External id": 5448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011601.434, "dur": 3.787,
    "args": {
      "External id": 5449,"Record function id": 0, "Ev Idx": 4936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011602.659, "dur": 1.837,
    "args": {
      "External id": 5450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011603.242, "dur": 0.925,
    "args": {
      "External id": 5451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011603.603, "dur": 0.418,
    "args": {
      "External id": 5452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011608.393, "dur": 4.321,
    "args": {
      "External id": 5453,"Record function id": 0, "Ev Idx": 4940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011609.570, "dur": 2.435,
    "args": {
      "External id": 5454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011610.174, "dur": 1.502,
    "args": {
      "External id": 5455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011610.550, "dur": 0.987,
    "args": {
      "External id": 5456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011615.877, "dur": 3.821,
    "args": {
      "External id": 5457,"Record function id": 0, "Ev Idx": 4944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011617.089, "dur": 1.874,
    "args": {
      "External id": 5458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011617.670, "dur": 0.966,
    "args": {
      "External id": 5459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011618.050, "dur": 0.444,
    "args": {
      "External id": 5460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011622.888, "dur": 4.144,
    "args": {
      "External id": 5461,"Record function id": 0, "Ev Idx": 4948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011624.119, "dur": 2.179,
    "args": {
      "External id": 5462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011624.832, "dur": 1.016,
    "args": {
      "External id": 5463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011625.236, "dur": 0.470,
    "args": {
      "External id": 5464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011630.236, "dur": 4.578,
    "args": {
      "External id": 5465,"Record function id": 0, "Ev Idx": 4952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011631.445, "dur": 2.623,
    "args": {
      "External id": 5466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011632.164, "dur": 1.469,
    "args": {
      "External id": 5467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011632.582, "dur": 0.909,
    "args": {
      "External id": 5468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011638.103, "dur": 3.939,
    "args": {
      "External id": 5469,"Record function id": 0, "Ev Idx": 4956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011639.417, "dur": 1.913,
    "args": {
      "External id": 5470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011640.013, "dur": 0.984,
    "args": {
      "External id": 5471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011640.378, "dur": 0.478,
    "args": {
      "External id": 5472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011645.233, "dur": 4.211,
    "args": {
      "External id": 5473,"Record function id": 0, "Ev Idx": 4960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011646.534, "dur": 2.158,
    "args": {
      "External id": 5474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011647.266, "dur": 0.998,
    "args": {
      "External id": 5475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011647.677, "dur": 0.449,
    "args": {
      "External id": 5476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011652.694, "dur": 3.952,
    "args": {
      "External id": 5477,"Record function id": 0, "Ev Idx": 4964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011653.966, "dur": 1.982,
    "args": {
      "External id": 5478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011654.653, "dur": 0.974,
    "args": {
      "External id": 5479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011655.027, "dur": 0.459,
    "args": {
      "External id": 5480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011659.846, "dur": 3.895,
    "args": {
      "External id": 5481,"Record function id": 0, "Ev Idx": 4968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011661.141, "dur": 1.878,
    "args": {
      "External id": 5482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011661.714, "dur": 0.969,
    "args": {
      "External id": 5483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011662.075, "dur": 0.466,
    "args": {
      "External id": 5484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 4971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011666.939, "dur": 4.063,
    "args": {
      "External id": 5485,"Record function id": 0, "Ev Idx": 4972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011668.125, "dur": 2.167,
    "args": {
      "External id": 5486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011668.735, "dur": 1.228,
    "args": {
      "External id": 5487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011669.109, "dur": 0.714,
    "args": {
      "External id": 5488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 4975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011674.187, "dur": 4.308,
    "args": {
      "External id": 5489,"Record function id": 0, "Ev Idx": 4976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011675.439, "dur": 2.284,
    "args": {
      "External id": 5490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011676.017, "dur": 1.352,
    "args": {
      "External id": 5491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011676.393, "dur": 0.833,
    "args": {
      "External id": 5492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 4979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011681.742, "dur": 3.793,
    "args": {
      "External id": 5493,"Record function id": 0, "Ev Idx": 4980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011682.942, "dur": 1.881,
    "args": {
      "External id": 5494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011683.528, "dur": 0.962,
    "args": {
      "External id": 5495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011683.901, "dur": 0.449,
    "args": {
      "External id": 5496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011688.701, "dur": 4.077,
    "args": {
      "External id": 5497,"Record function id": 0, "Ev Idx": 4984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011689.917, "dur": 2.136,
    "args": {
      "External id": 5498,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011690.513, "dur": 1.212,
    "args": {
      "External id": 5499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011690.901, "dur": 0.682,
    "args": {
      "External id": 5500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 4987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011696.000, "dur": 3.992,
    "args": {
      "External id": 5501,"Record function id": 0, "Ev Idx": 4988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011697.224, "dur": 2.061,
    "args": {
      "External id": 5502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011697.790, "dur": 1.169,
    "args": {
      "External id": 5503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011698.151, "dur": 0.669,
    "args": {
      "External id": 5504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 4991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011703.199, "dur": 4.002,
    "args": {
      "External id": 5505,"Record function id": 0, "Ev Idx": 4992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011704.450, "dur": 2.020,
    "args": {
      "External id": 5506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011705.059, "dur": 1.087,
    "args": {
      "External id": 5507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011705.476, "dur": 0.525,
    "args": {
      "External id": 5508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 4995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011710.367, "dur": 4.239,
    "args": {
      "External id": 5509,"Record function id": 0, "Ev Idx": 4996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011711.562, "dur": 2.322,
    "args": {
      "External id": 5510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011712.157, "dur": 1.402,
    "args": {
      "External id": 5511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011712.527, "dur": 0.887,
    "args": {
      "External id": 5512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 4999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011717.756, "dur": 4.661,
    "args": {
      "External id": 5513,"Record function id": 0, "Ev Idx": 5000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011718.986, "dur": 2.660,
    "args": {
      "External id": 5514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011719.755, "dur": 1.452,
    "args": {
      "External id": 5515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011720.164, "dur": 0.899,
    "args": {
      "External id": 5516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011725.655, "dur": 4.256,
    "args": {
      "External id": 5517,"Record function id": 0, "Ev Idx": 5004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011726.855, "dur": 2.332,
    "args": {
      "External id": 5518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011727.590, "dur": 1.170,
    "args": {
      "External id": 5519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011727.979, "dur": 0.638,
    "args": {
      "External id": 5520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011733.109, "dur": 4.697,
    "args": {
      "External id": 5521,"Record function id": 0, "Ev Idx": 5008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011734.336, "dur": 2.717,
    "args": {
      "External id": 5522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011734.980, "dur": 1.731,
    "args": {
      "External id": 5523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011735.480, "dur": 1.053,
    "args": {
      "External id": 5524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011741.038, "dur": 3.896,
    "args": {
      "External id": 5525,"Record function id": 0, "Ev Idx": 5012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011742.241, "dur": 1.964,
    "args": {
      "External id": 5526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011742.859, "dur": 1.027,
    "args": {
      "External id": 5527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011743.234, "dur": 0.513,
    "args": {
      "External id": 5528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011748.080, "dur": 4.122,
    "args": {
      "External id": 5529,"Record function id": 0, "Ev Idx": 5016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011749.325, "dur": 2.147,
    "args": {
      "External id": 5530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011749.935, "dur": 1.202,
    "args": {
      "External id": 5531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011750.335, "dur": 0.664,
    "args": {
      "External id": 5532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011755.442, "dur": 3.823,
    "args": {
      "External id": 5533,"Record function id": 0, "Ev Idx": 5020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011756.662, "dur": 1.890,
    "args": {
      "External id": 5534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011757.230, "dur": 0.987,
    "args": {
      "External id": 5535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011757.616, "dur": 0.459,
    "args": {
      "External id": 5536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011762.451, "dur": 3.882,
    "args": {
      "External id": 5537,"Record function id": 0, "Ev Idx": 5024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011763.644, "dur": 1.965,
    "args": {
      "External id": 5538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011764.238, "dur": 1.037,
    "args": {
      "External id": 5539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011764.623, "dur": 0.510,
    "args": {
      "External id": 5540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011769.559, "dur": 4.729,
    "args": {
      "External id": 5541,"Record function id": 0, "Ev Idx": 5028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011770.738, "dur": 2.840,
    "args": {
      "External id": 5542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011771.310, "dur": 1.935,
    "args": {
      "External id": 5543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011771.720, "dur": 1.384,
    "args": {
      "External id": 5544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011777.460, "dur": 4.279,
    "args": {
      "External id": 5545,"Record function id": 0, "Ev Idx": 5032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011778.705, "dur": 2.306,
    "args": {
      "External id": 5546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011779.403, "dur": 1.282,
    "args": {
      "External id": 5547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011779.782, "dur": 0.754,
    "args": {
      "External id": 5548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011784.910, "dur": 4.251,
    "args": {
      "External id": 5549,"Record function id": 0, "Ev Idx": 5036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011786.064, "dur": 2.385,
    "args": {
      "External id": 5550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011786.746, "dur": 1.375,
    "args": {
      "External id": 5551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011787.119, "dur": 0.860,
    "args": {
      "External id": 5552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011792.341, "dur": 4.055,
    "args": {
      "External id": 5553,"Record function id": 0, "Ev Idx": 5040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011793.576, "dur": 2.097,
    "args": {
      "External id": 5554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011794.344, "dur": 0.986,
    "args": {
      "External id": 5555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011794.711, "dur": 0.477,
    "args": {
      "External id": 5556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011799.662, "dur": 3.698,
    "args": {
      "External id": 5557,"Record function id": 0, "Ev Idx": 5044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011800.844, "dur": 1.809,
    "args": {
      "External id": 5558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011801.470, "dur": 0.870,
    "args": {
      "External id": 5559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011801.846, "dur": 0.354,
    "args": {
      "External id": 5560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011806.526, "dur": 4.565,
    "args": {
      "External id": 5561,"Record function id": 0, "Ev Idx": 5048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011808.292, "dur": 2.072,
    "args": {
      "External id": 5562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011808.924, "dur": 1.091,
    "args": {
      "External id": 5563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011809.322, "dur": 0.550,
    "args": {
      "External id": 5564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011814.566, "dur": 3.965,
    "args": {
      "External id": 5565,"Record function id": 0, "Ev Idx": 5052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011815.871, "dur": 1.954,
    "args": {
      "External id": 5566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011816.461, "dur": 1.038,
    "args": {
      "External id": 5567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011816.841, "dur": 0.514,
    "args": {
      "External id": 5568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011821.709, "dur": 4.061,
    "args": {
      "External id": 5569,"Record function id": 0, "Ev Idx": 5056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011822.938, "dur": 2.105,
    "args": {
      "External id": 5570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011823.565, "dur": 1.150,
    "args": {
      "External id": 5571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011823.942, "dur": 0.627,
    "args": {
      "External id": 5572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011828.974, "dur": 4.079,
    "args": {
      "External id": 5573,"Record function id": 0, "Ev Idx": 5060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011830.165, "dur": 2.182,
    "args": {
      "External id": 5574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011830.755, "dur": 1.262,
    "args": {
      "External id": 5575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011831.130, "dur": 0.743,
    "args": {
      "External id": 5576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011836.263, "dur": 3.826,
    "args": {
      "External id": 5577,"Record function id": 0, "Ev Idx": 5064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011837.499, "dur": 1.865,
    "args": {
      "External id": 5578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 5065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011838.085, "dur": 0.954,
    "args": {
      "External id": 5579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 5066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011838.453, "dur": 0.441,
    "args": {
      "External id": 5580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 5067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011843.262, "dur": 3.858,
    "args": {
      "External id": 5581,"Record function id": 0, "Ev Idx": 5068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011844.586, "dur": 1.818,
    "args": {
      "External id": 5582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011845.184, "dur": 0.896,
    "args": {
      "External id": 5583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011845.570, "dur": 0.368,
    "args": {
      "External id": 5584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011850.348, "dur": 4.002,
    "args": {
      "External id": 5585,"Record function id": 0, "Ev Idx": 5072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011851.521, "dur": 2.067,
    "args": {
      "External id": 5586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011852.266, "dur": 0.845,
    "args": {
      "External id": 5587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011852.629, "dur": 0.344,
    "args": {
      "External id": 5588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011857.594, "dur": 4.133,
    "args": {
      "External id": 5589,"Record function id": 0, "Ev Idx": 5076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011858.800, "dur": 2.196,
    "args": {
      "External id": 5590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011859.580, "dur": 0.992,
    "args": {
      "External id": 5591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011859.945, "dur": 0.485,
    "args": {
      "External id": 5592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011864.946, "dur": 3.880,
    "args": {
      "External id": 5593,"Record function id": 0, "Ev Idx": 5080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011866.123, "dur": 1.963,
    "args": {
      "External id": 5594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011866.736, "dur": 1.026,
    "args": {
      "External id": 5595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011867.110, "dur": 0.517,
    "args": {
      "External id": 5596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011872.269, "dur": 3.722,
    "args": {
      "External id": 5597,"Record function id": 0, "Ev Idx": 5084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011873.465, "dur": 1.808,
    "args": {
      "External id": 5598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011874.043, "dur": 0.900,
    "args": {
      "External id": 5599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011874.448, "dur": 0.365,
    "args": {
      "External id": 5600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011879.691, "dur": 4.248,
    "args": {
      "External id": 5601,"Record function id": 0, "Ev Idx": 5088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011881.009, "dur": 2.161,
    "args": {
      "External id": 5602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011881.663, "dur": 1.149,
    "args": {
      "External id": 5603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011882.084, "dur": 0.539,
    "args": {
      "External id": 5604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011887.181, "dur": 4.071,
    "args": {
      "External id": 5605,"Record function id": 0, "Ev Idx": 5092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011888.497, "dur": 2.042,
    "args": {
      "External id": 5606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011889.140, "dur": 1.076,
    "args": {
      "External id": 5607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011889.513, "dur": 0.559,
    "args": {
      "External id": 5608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011894.512, "dur": 3.831,
    "args": {
      "External id": 5609,"Record function id": 0, "Ev Idx": 5096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011895.837, "dur": 1.782,
    "args": {
      "External id": 5610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011896.424, "dur": 0.861,
    "args": {
      "External id": 5611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011896.795, "dur": 0.350,
    "args": {
      "External id": 5612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011901.567, "dur": 3.850,
    "args": {
      "External id": 5613,"Record function id": 0, "Ev Idx": 5100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011902.791, "dur": 1.904,
    "args": {
      "External id": 5614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011903.386, "dur": 0.982,
    "args": {
      "External id": 5615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011903.772, "dur": 0.452,
    "args": {
      "External id": 5616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011908.711, "dur": 4.278,
    "args": {
      "External id": 5617,"Record function id": 0, "Ev Idx": 5104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011909.939, "dur": 2.307,
    "args": {
      "External id": 5618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011910.518, "dur": 1.398,
    "args": {
      "External id": 5619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011910.885, "dur": 0.898,
    "args": {
      "External id": 5620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011916.244, "dur": 4.070,
    "args": {
      "External id": 5621,"Record function id": 0, "Ev Idx": 5108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011917.452, "dur": 2.152,
    "args": {
      "External id": 5622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011918.156, "dur": 1.126,
    "args": {
      "External id": 5623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011918.634, "dur": 0.522,
    "args": {
      "External id": 5624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011923.656, "dur": 3.967,
    "args": {
      "External id": 5625,"Record function id": 0, "Ev Idx": 5112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011925.022, "dur": 1.874,
    "args": {
      "External id": 5626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011925.599, "dur": 0.963,
    "args": {
      "External id": 5627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011925.967, "dur": 0.454,
    "args": {
      "External id": 5628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011931.129, "dur": 11.296,
    "args": {
      "External id": 5629,"Record function id": 0, "Ev Idx": 5116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508011936.709, "dur": 4.801,
    "args": {
      "External id": 5630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011937.934, "dur": 3.054,
    "args": {
      "External id": 5631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508011938.491, "dur": 2.209,
    "args": {
      "External id": 5632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012307.693, "dur": 19.303,
    "args": {
      "External id": 5633,"Record function id": 0, "Ev Idx": 5120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012319.317, "dur": 6.129,
    "args": {
      "External id": 5634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012321.925, "dur": 2.567,
    "args": {
      "External id": 5635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012322.648, "dur": 1.472,
    "args": {
      "External id": 5636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012335.548, "dur": 4.402,
    "args": {
      "External id": 5637,"Record function id": 0, "Ev Idx": 5124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012337.091, "dur": 2.070,
    "args": {
      "External id": 5638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012337.813, "dur": 1.015,
    "args": {
      "External id": 5639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012338.226, "dur": 0.461,
    "args": {
      "External id": 5640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012343.505, "dur": 3.761,
    "args": {
      "External id": 5641,"Record function id": 0, "Ev Idx": 5128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012344.835, "dur": 1.730,
    "args": {
      "External id": 5642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012345.423, "dur": 0.842,
    "args": {
      "External id": 5643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012345.758, "dur": 0.377,
    "args": {
      "External id": 5644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012350.603, "dur": 3.891,
    "args": {
      "External id": 5645,"Record function id": 0, "Ev Idx": 5132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012351.910, "dur": 1.875,
    "args": {
      "External id": 5646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012352.542, "dur": 0.923,
    "args": {
      "External id": 5647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012352.927, "dur": 0.394,
    "args": {
      "External id": 5648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012357.754, "dur": 3.771,
    "args": {
      "External id": 5649,"Record function id": 0, "Ev Idx": 5136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012359.004, "dur": 1.769,
    "args": {
      "External id": 5650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012359.568, "dur": 0.906,
    "args": {
      "External id": 5651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012359.929, "dur": 0.396,
    "args": {
      "External id": 5652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012364.874, "dur": 4.672,
    "args": {
      "External id": 5653,"Record function id": 0, "Ev Idx": 5140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012366.153, "dur": 2.705,
    "args": {
      "External id": 5654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012367.683, "dur": 0.861,
    "args": {
      "External id": 5655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012368.065, "dur": 0.352,
    "args": {
      "External id": 5656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012372.687, "dur": 3.768,
    "args": {
      "External id": 5657,"Record function id": 0, "Ev Idx": 5144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012373.970, "dur": 1.780,
    "args": {
      "External id": 5658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012374.539, "dur": 0.913,
    "args": {
      "External id": 5659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012374.928, "dur": 0.377,
    "args": {
      "External id": 5660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012379.854, "dur": 3.885,
    "args": {
      "External id": 5661,"Record function id": 0, "Ev Idx": 5148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012381.325, "dur": 1.696,
    "args": {
      "External id": 5662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012381.924, "dur": 0.804,
    "args": {
      "External id": 5663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012382.255, "dur": 0.345,
    "args": {
      "External id": 5664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012386.932, "dur": 3.814,
    "args": {
      "External id": 5665,"Record function id": 0, "Ev Idx": 5152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012388.183, "dur": 1.863,
    "args": {
      "External id": 5666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012388.751, "dur": 0.997,
    "args": {
      "External id": 5667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012389.125, "dur": 0.486,
    "args": {
      "External id": 5668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012393.925, "dur": 3.821,
    "args": {
      "External id": 5669,"Record function id": 0, "Ev Idx": 5156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012395.183, "dur": 1.863,
    "args": {
      "External id": 5670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012395.814, "dur": 0.934,
    "args": {
      "External id": 5671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012396.269, "dur": 0.333,
    "args": {
      "External id": 5672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012400.994, "dur": 3.958,
    "args": {
      "External id": 5673,"Record function id": 0, "Ev Idx": 5160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012402.272, "dur": 1.940,
    "args": {
      "External id": 5674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012402.993, "dur": 0.772,
    "args": {
      "External id": 5675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012403.344, "dur": 0.290,
    "args": {
      "External id": 5676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012408.273, "dur": 3.833,
    "args": {
      "External id": 5677,"Record function id": 0, "Ev Idx": 5164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012409.528, "dur": 1.856,
    "args": {
      "External id": 5678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012410.174, "dur": 0.806,
    "args": {
      "External id": 5679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012410.509, "dur": 0.331,
    "args": {
      "External id": 5680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012415.295, "dur": 4.593,
    "args": {
      "External id": 5681,"Record function id": 0, "Ev Idx": 5168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012417.380, "dur": 1.767,
    "args": {
      "External id": 5682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012418.048, "dur": 0.791,
    "args": {
      "External id": 5683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012418.416, "dur": 0.293,
    "args": {
      "External id": 5684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012423.033, "dur": 3.683,
    "args": {
      "External id": 5685,"Record function id": 0, "Ev Idx": 5172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012424.338, "dur": 1.678,
    "args": {
      "External id": 5686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012424.865, "dur": 0.856,
    "args": {
      "External id": 5687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012425.287, "dur": 0.303,
    "args": {
      "External id": 5688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012429.889, "dur": 3.591,
    "args": {
      "External id": 5689,"Record function id": 0, "Ev Idx": 5176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012431.111, "dur": 1.638,
    "args": {
      "External id": 5690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012431.650, "dur": 0.804,
    "args": {
      "External id": 5691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012432.010, "dur": 0.311,
    "args": {
      "External id": 5692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012436.679, "dur": 3.660,
    "args": {
      "External id": 5693,"Record function id": 0, "Ev Idx": 5180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012437.922, "dur": 1.710,
    "args": {
      "External id": 5694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012438.447, "dur": 0.893,
    "args": {
      "External id": 5695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012438.813, "dur": 0.388,
    "args": {
      "External id": 5696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012443.517, "dur": 3.761,
    "args": {
      "External id": 5697,"Record function id": 0, "Ev Idx": 5184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012444.756, "dur": 1.825,
    "args": {
      "External id": 5698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012445.305, "dur": 0.969,
    "args": {
      "External id": 5699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012445.643, "dur": 0.503,
    "args": {
      "External id": 5700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012450.397, "dur": 3.758,
    "args": {
      "External id": 5701,"Record function id": 0, "Ev Idx": 5188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012451.754, "dur": 1.703,
    "args": {
      "External id": 5702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012452.317, "dur": 0.841,
    "args": {
      "External id": 5703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012452.677, "dur": 0.343,
    "args": {
      "External id": 5704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012457.381, "dur": 3.680,
    "args": {
      "External id": 5705,"Record function id": 0, "Ev Idx": 5192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012458.686, "dur": 1.694,
    "args": {
      "External id": 5706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012459.237, "dur": 0.831,
    "args": {
      "External id": 5707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012459.584, "dur": 0.343,
    "args": {
      "External id": 5708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012465.252, "dur": 3.648,
    "args": {
      "External id": 5709,"Record function id": 0, "Ev Idx": 5196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012466.499, "dur": 1.718,
    "args": {
      "External id": 5710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012467.058, "dur": 0.849,
    "args": {
      "External id": 5711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012467.420, "dur": 0.345,
    "args": {
      "External id": 5712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012472.104, "dur": 3.698,
    "args": {
      "External id": 5713,"Record function id": 0, "Ev Idx": 5200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012473.457, "dur": 1.669,
    "args": {
      "External id": 5714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012474.030, "dur": 0.785,
    "args": {
      "External id": 5715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012474.375, "dur": 0.313,
    "args": {
      "External id": 5716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012478.970, "dur": 3.702,
    "args": {
      "External id": 5717,"Record function id": 0, "Ev Idx": 5204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012480.195, "dur": 1.803,
    "args": {
      "External id": 5718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012480.742, "dur": 0.958,
    "args": {
      "External id": 5719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012481.083, "dur": 0.487,
    "args": {
      "External id": 5720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012485.918, "dur": 4.004,
    "args": {
      "External id": 5721,"Record function id": 0, "Ev Idx": 5208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012487.285, "dur": 1.921,
    "args": {
      "External id": 5722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012487.948, "dur": 0.832,
    "args": {
      "External id": 5723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012488.300, "dur": 0.338,
    "args": {
      "External id": 5724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012493.177, "dur": 5.690,
    "args": {
      "External id": 5725,"Record function id": 0, "Ev Idx": 5212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012494.387, "dur": 3.780,
    "args": {
      "External id": 5726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012496.981, "dur": 0.881,
    "args": {
      "External id": 5727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012497.331, "dur": 0.400,
    "args": {
      "External id": 5728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012502.133, "dur": 3.744,
    "args": {
      "External id": 5729,"Record function id": 0, "Ev Idx": 5216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012503.360, "dur": 1.823,
    "args": {
      "External id": 5730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012503.916, "dur": 0.962,
    "args": {
      "External id": 5731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012504.259, "dur": 0.489,
    "args": {
      "External id": 5732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012509.177, "dur": 4.573,
    "args": {
      "External id": 5733,"Record function id": 0, "Ev Idx": 5220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012510.417, "dur": 2.618,
    "args": {
      "External id": 5734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012510.958, "dur": 1.751,
    "args": {
      "External id": 5735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012512.208, "dur": 0.357,
    "args": {
      "External id": 5736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012516.964, "dur": 3.682,
    "args": {
      "External id": 5737,"Record function id": 0, "Ev Idx": 5224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012518.187, "dur": 1.761,
    "args": {
      "External id": 5738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012518.725, "dur": 0.922,
    "args": {
      "External id": 5739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012519.073, "dur": 0.436,
    "args": {
      "External id": 5740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012523.837, "dur": 3.774,
    "args": {
      "External id": 5741,"Record function id": 0, "Ev Idx": 5228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012525.236, "dur": 1.676,
    "args": {
      "External id": 5742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012525.770, "dur": 0.841,
    "args": {
      "External id": 5743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012526.141, "dur": 0.338,
    "args": {
      "External id": 5744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012530.847, "dur": 3.619,
    "args": {
      "External id": 5745,"Record function id": 0, "Ev Idx": 5232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012532.123, "dur": 1.650,
    "args": {
      "External id": 5746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012532.651, "dur": 0.814,
    "args": {
      "External id": 5747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012532.983, "dur": 0.336,
    "args": {
      "External id": 5748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012537.609, "dur": 3.807,
    "args": {
      "External id": 5749,"Record function id": 0, "Ev Idx": 5236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012538.815, "dur": 1.902,
    "args": {
      "External id": 5750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012539.395, "dur": 1.020,
    "args": {
      "External id": 5751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012539.786, "dur": 0.492,
    "args": {
      "External id": 5752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012544.590, "dur": 3.682,
    "args": {
      "External id": 5753,"Record function id": 0, "Ev Idx": 5240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012545.897, "dur": 1.661,
    "args": {
      "External id": 5754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012546.448, "dur": 0.810,
    "args": {
      "External id": 5755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012546.805, "dur": 0.321,
    "args": {
      "External id": 5756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012551.502, "dur": 3.796,
    "args": {
      "External id": 5757,"Record function id": 0, "Ev Idx": 5244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012552.729, "dur": 1.857,
    "args": {
      "External id": 5758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012553.380, "dur": 0.790,
    "args": {
      "External id": 5759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012553.741, "dur": 0.297,
    "args": {
      "External id": 5760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012558.520, "dur": 4.970,
    "args": {
      "External id": 5761,"Record function id": 0, "Ev Idx": 5248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012559.773, "dur": 3.021,
    "args": {
      "External id": 5762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012561.284, "dur": 1.053,
    "args": {
      "External id": 5763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012561.636, "dur": 0.560,
    "args": {
      "External id": 5764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012566.595, "dur": 3.584,
    "args": {
      "External id": 5765,"Record function id": 0, "Ev Idx": 5252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012567.858, "dur": 1.634,
    "args": {
      "External id": 5766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012568.410, "dur": 0.780,
    "args": {
      "External id": 5767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012568.762, "dur": 0.298,
    "args": {
      "External id": 5768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012573.295, "dur": 3.592,
    "args": {
      "External id": 5769,"Record function id": 0, "Ev Idx": 5256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012574.547, "dur": 1.665,
    "args": {
      "External id": 5770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012575.107, "dur": 0.810,
    "args": {
      "External id": 5771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012575.435, "dur": 0.341,
    "args": {
      "External id": 5772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012580.047, "dur": 3.689,
    "args": {
      "External id": 5773,"Record function id": 0, "Ev Idx": 5260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012581.284, "dur": 1.735,
    "args": {
      "External id": 5774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012581.899, "dur": 0.822,
    "args": {
      "External id": 5775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012582.282, "dur": 0.303,
    "args": {
      "External id": 5776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012586.881, "dur": 3.583,
    "args": {
      "External id": 5777,"Record function id": 0, "Ev Idx": 5264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012588.099, "dur": 1.689,
    "args": {
      "External id": 5778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012588.648, "dur": 0.838,
    "args": {
      "External id": 5779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012588.991, "dur": 0.355,
    "args": {
      "External id": 5780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012593.598, "dur": 3.723,
    "args": {
      "External id": 5781,"Record function id": 0, "Ev Idx": 5268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012594.847, "dur": 1.788,
    "args": {
      "External id": 5782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012595.372, "dur": 0.964,
    "args": {
      "External id": 5783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012595.735, "dur": 0.475,
    "args": {
      "External id": 5784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012600.426, "dur": 3.578,
    "args": {
      "External id": 5785,"Record function id": 0, "Ev Idx": 5272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012601.633, "dur": 1.696,
    "args": {
      "External id": 5786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012602.199, "dur": 0.833,
    "args": {
      "External id": 5787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012602.554, "dur": 0.342,
    "args": {
      "External id": 5788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012607.169, "dur": 4.522,
    "args": {
      "External id": 5789,"Record function id": 0, "Ev Idx": 5276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012609.269, "dur": 1.742,
    "args": {
      "External id": 5790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012609.809, "dur": 0.902,
    "args": {
      "External id": 5791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012610.173, "dur": 0.403,
    "args": {
      "External id": 5792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012615.018, "dur": 5.580,
    "args": {
      "External id": 5793,"Record function id": 0, "Ev Idx": 5280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012618.116, "dur": 1.797,
    "args": {
      "External id": 5794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012618.668, "dur": 0.875,
    "args": {
      "External id": 5795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012619.015, "dur": 0.403,
    "args": {
      "External id": 5796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012623.865, "dur": 3.661,
    "args": {
      "External id": 5797,"Record function id": 0, "Ev Idx": 5284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012625.180, "dur": 1.646,
    "args": {
      "External id": 5798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012625.721, "dur": 0.801,
    "args": {
      "External id": 5799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012626.076, "dur": 0.316,
    "args": {
      "External id": 5800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012630.682, "dur": 3.956,
    "args": {
      "External id": 5801,"Record function id": 0, "Ev Idx": 5288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012632.032, "dur": 1.884,
    "args": {
      "External id": 5802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012632.702, "dur": 0.787,
    "args": {
      "External id": 5803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012633.064, "dur": 0.302,
    "args": {
      "External id": 5804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012637.806, "dur": 3.860,
    "args": {
      "External id": 5805,"Record function id": 0, "Ev Idx": 5292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012639.081, "dur": 1.867,
    "args": {
      "External id": 5806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012639.756, "dur": 0.755,
    "args": {
      "External id": 5807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012640.088, "dur": 0.288,
    "args": {
      "External id": 5808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012644.833, "dur": 3.649,
    "args": {
      "External id": 5809,"Record function id": 0, "Ev Idx": 5296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012646.093, "dur": 1.669,
    "args": {
      "External id": 5810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012646.638, "dur": 0.832,
    "args": {
      "External id": 5811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012646.993, "dur": 0.345,
    "args": {
      "External id": 5812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012651.693, "dur": 3.571,
    "args": {
      "External id": 5813,"Record function id": 0, "Ev Idx": 5300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012652.927, "dur": 1.645,
    "args": {
      "External id": 5814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012653.465, "dur": 0.810,
    "args": {
      "External id": 5815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012653.827, "dur": 0.325,
    "args": {
      "External id": 5816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012659.873, "dur": 3.758,
    "args": {
      "External id": 5817,"Record function id": 0, "Ev Idx": 5304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012661.139, "dur": 1.791,
    "args": {
      "External id": 5818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012661.708, "dur": 0.909,
    "args": {
      "External id": 5819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012662.165, "dur": 0.321,
    "args": {
      "External id": 5820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012666.811, "dur": 3.635,
    "args": {
      "External id": 5821,"Record function id": 0, "Ev Idx": 5308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012668.090, "dur": 1.657,
    "args": {
      "External id": 5822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012668.635, "dur": 0.806,
    "args": {
      "External id": 5823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012668.990, "dur": 0.323,
    "args": {
      "External id": 5824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012673.549, "dur": 3.880,
    "args": {
      "External id": 5825,"Record function id": 0, "Ev Idx": 5312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012674.761, "dur": 1.958,
    "args": {
      "External id": 5826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012675.318, "dur": 1.099,
    "args": {
      "External id": 5827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012675.659, "dur": 0.630,
    "args": {
      "External id": 5828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012680.507, "dur": 3.904,
    "args": {
      "External id": 5829,"Record function id": 0, "Ev Idx": 5316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012681.735, "dur": 1.968,
    "args": {
      "External id": 5830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012682.327, "dur": 1.079,
    "args": {
      "External id": 5831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012682.684, "dur": 0.588,
    "args": {
      "External id": 5832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012687.486, "dur": 4.133,
    "args": {
      "External id": 5833,"Record function id": 0, "Ev Idx": 5320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012688.710, "dur": 2.186,
    "args": {
      "External id": 5834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012689.328, "dur": 1.277,
    "args": {
      "External id": 5835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012689.698, "dur": 0.776,
    "args": {
      "External id": 5836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012694.690, "dur": 4.398,
    "args": {
      "External id": 5837,"Record function id": 0, "Ev Idx": 5324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012695.886, "dur": 2.506,
    "args": {
      "External id": 5838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012696.417, "dur": 1.673,
    "args": {
      "External id": 5839,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012696.752, "dur": 1.208,
    "args": {
      "External id": 5840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012702.151, "dur": 4.246,
    "args": {
      "External id": 5841,"Record function id": 0, "Ev Idx": 5328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012703.412, "dur": 2.255,
    "args": {
      "External id": 5842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012704.102, "dur": 1.115,
    "args": {
      "External id": 5843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012704.442, "dur": 0.648,
    "args": {
      "External id": 5844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012709.499, "dur": 4.303,
    "args": {
      "External id": 5845,"Record function id": 0, "Ev Idx": 5332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012710.682, "dur": 2.412,
    "args": {
      "External id": 5846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012711.343, "dur": 1.336,
    "args": {
      "External id": 5847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012711.682, "dur": 0.867,
    "args": {
      "External id": 5848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012716.961, "dur": 3.911,
    "args": {
      "External id": 5849,"Record function id": 0, "Ev Idx": 5336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012718.214, "dur": 1.967,
    "args": {
      "External id": 5850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012718.759, "dur": 1.115,
    "args": {
      "External id": 5851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012719.159, "dur": 0.575,
    "args": {
      "External id": 5852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012723.988, "dur": 3.667,
    "args": {
      "External id": 5853,"Record function id": 0, "Ev Idx": 5340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012725.211, "dur": 1.724,
    "args": {
      "External id": 5854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012725.783, "dur": 0.848,
    "args": {
      "External id": 5855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012726.148, "dur": 0.357,
    "args": {
      "External id": 5856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012730.710, "dur": 4.104,
    "args": {
      "External id": 5857,"Record function id": 0, "Ev Idx": 5344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012732.050, "dur": 2.083,
    "args": {
      "External id": 5858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012732.581, "dur": 1.249,
    "args": {
      "External id": 5859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012732.952, "dur": 0.748,
    "args": {
      "External id": 5860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012737.879, "dur": 3.744,
    "args": {
      "External id": 5861,"Record function id": 0, "Ev Idx": 5348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012739.045, "dur": 1.879,
    "args": {
      "External id": 5862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012739.613, "dur": 1.012,
    "args": {
      "External id": 5863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012739.977, "dur": 0.509,
    "args": {
      "External id": 5864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012744.714, "dur": 3.886,
    "args": {
      "External id": 5865,"Record function id": 0, "Ev Idx": 5352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012745.950, "dur": 1.946,
    "args": {
      "External id": 5866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012746.521, "dur": 1.040,
    "args": {
      "External id": 5867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012746.909, "dur": 0.522,
    "args": {
      "External id": 5868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012751.736, "dur": 3.840,
    "args": {
      "External id": 5869,"Record function id": 0, "Ev Idx": 5356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012752.988, "dur": 1.901,
    "args": {
      "External id": 5870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012753.541, "dur": 1.045,
    "args": {
      "External id": 5871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012753.920, "dur": 0.533,
    "args": {
      "External id": 5872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012758.651, "dur": 3.696,
    "args": {
      "External id": 5873,"Record function id": 0, "Ev Idx": 5360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012759.885, "dur": 1.780,
    "args": {
      "External id": 5874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012760.426, "dur": 0.941,
    "args": {
      "External id": 5875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012760.792, "dur": 0.443,
    "args": {
      "External id": 5876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012765.440, "dur": 3.764,
    "args": {
      "External id": 5877,"Record function id": 0, "Ev Idx": 5364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012766.592, "dur": 1.939,
    "args": {
      "External id": 5878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012767.178, "dur": 1.046,
    "args": {
      "External id": 5879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012767.555, "dur": 0.532,
    "args": {
      "External id": 5880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012772.274, "dur": 4.254,
    "args": {
      "External id": 5881,"Record function id": 0, "Ev Idx": 5368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012773.486, "dur": 2.342,
    "args": {
      "External id": 5882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012774.212, "dur": 1.318,
    "args": {
      "External id": 5883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012774.569, "dur": 0.829,
    "args": {
      "External id": 5884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012779.562, "dur": 3.776,
    "args": {
      "External id": 5885,"Record function id": 0, "Ev Idx": 5372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012780.726, "dur": 1.919,
    "args": {
      "External id": 5886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012781.272, "dur": 1.076,
    "args": {
      "External id": 5887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012781.680, "dur": 0.540,
    "args": {
      "External id": 5888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012786.403, "dur": 3.699,
    "args": {
      "External id": 5889,"Record function id": 0, "Ev Idx": 5376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012787.662, "dur": 1.740,
    "args": {
      "External id": 5890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012788.204, "dur": 0.900,
    "args": {
      "External id": 5891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012788.587, "dur": 0.389,
    "args": {
      "External id": 5892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012793.192, "dur": 3.759,
    "args": {
      "External id": 5893,"Record function id": 0, "Ev Idx": 5380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012794.567, "dur": 1.686,
    "args": {
      "External id": 5894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012795.109, "dur": 0.839,
    "args": {
      "External id": 5895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012795.472, "dur": 0.350,
    "args": {
      "External id": 5896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012800.031, "dur": 3.624,
    "args": {
      "External id": 5897,"Record function id": 0, "Ev Idx": 5384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012801.259, "dur": 1.698,
    "args": {
      "External id": 5898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012801.798, "dur": 0.855,
    "args": {
      "External id": 5899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012802.156, "dur": 0.367,
    "args": {
      "External id": 5900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012806.744, "dur": 3.947,
    "args": {
      "External id": 5901,"Record function id": 0, "Ev Idx": 5388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012807.986, "dur": 1.986,
    "args": {
      "External id": 5902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012808.552, "dur": 1.119,
    "args": {
      "External id": 5903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012808.913, "dur": 0.625,
    "args": {
      "External id": 5904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012813.771, "dur": 3.872,
    "args": {
      "External id": 5905,"Record function id": 0, "Ev Idx": 5392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012815.006, "dur": 1.944,
    "args": {
      "External id": 5906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 5393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012815.537, "dur": 1.115,
    "args": {
      "External id": 5907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 5394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012815.877, "dur": 0.645,
    "args": {
      "External id": 5908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 5395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012820.669, "dur": 3.832,
    "args": {
      "External id": 5909,"Record function id": 0, "Ev Idx": 5396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012821.853, "dur": 1.957,
    "args": {
      "External id": 5910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012822.524, "dur": 0.855,
    "args": {
      "External id": 5911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012822.907, "dur": 0.344,
    "args": {
      "External id": 5912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012827.622, "dur": 3.899,
    "args": {
      "External id": 5913,"Record function id": 0, "Ev Idx": 5400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012828.791, "dur": 2.041,
    "args": {
      "External id": 5914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012829.543, "dur": 0.840,
    "args": {
      "External id": 5915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012829.881, "dur": 0.370,
    "args": {
      "External id": 5916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012834.535, "dur": 3.586,
    "args": {
      "External id": 5917,"Record function id": 0, "Ev Idx": 5404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012835.748, "dur": 1.690,
    "args": {
      "External id": 5918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012836.298, "dur": 0.837,
    "args": {
      "External id": 5919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012836.678, "dur": 0.323,
    "args": {
      "External id": 5920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012841.178, "dur": 3.787,
    "args": {
      "External id": 5921,"Record function id": 0, "Ev Idx": 5408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012842.383, "dur": 1.898,
    "args": {
      "External id": 5922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012842.949, "dur": 1.026,
    "args": {
      "External id": 5923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012843.362, "dur": 0.478,
    "args": {
      "External id": 5924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012848.014, "dur": 3.812,
    "args": {
      "External id": 5925,"Record function id": 0, "Ev Idx": 5412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012849.331, "dur": 1.817,
    "args": {
      "External id": 5926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012849.877, "dur": 0.969,
    "args": {
      "External id": 5927,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012850.257, "dur": 0.457,
    "args": {
      "External id": 5928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012854.826, "dur": 3.664,
    "args": {
      "External id": 5929,"Record function id": 0, "Ev Idx": 5416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012856.025, "dur": 1.776,
    "args": {
      "External id": 5930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012856.552, "dur": 0.948,
    "args": {
      "External id": 5931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012856.883, "dur": 0.485,
    "args": {
      "External id": 5932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012861.568, "dur": 3.722,
    "args": {
      "External id": 5933,"Record function id": 0, "Ev Idx": 5420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508012862.767, "dur": 1.799,
    "args": {
      "External id": 5934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012863.378, "dur": 0.880,
    "args": {
      "External id": 5935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508012863.755, "dur": 0.368,
    "args": {
      "External id": 5936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508012868.580, "dur": 169.922,
    "args": {
      "External id": 5937,"Record function id": 0, "Sequence number": 4696, "Fwd thread id": 1, "Ev Idx": 5424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508012870.340, "dur": 129.683,
    "args": {
      "External id": 5938,"Sequence number": 4696, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 5425
    }
  },
  {
    "ph": "f", "id": 14, "pid": 14565, "tid": 14640, "ts": 1290508012870.340,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290508012929.562, "dur": 60.184,
    "args": {
      "External id": 5939,"Record function id": 0, "scope": "user_scope", "Ev Idx": 5426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290508013008.585, "dur": 27.240,
    "args": {
      "External id": 5940,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 5427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508013045.358, "dur": 6.229,
    "args": {
      "External id": 5941,"Record function id": 0, "Ev Idx": 5428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508013047.191, "dur": 3.504,
    "args": {
      "External id": 5942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 5429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508013048.350, "dur": 1.939,
    "args": {
      "External id": 5943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 5430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508013048.991, "dur": 1.090,
    "args": {
      "External id": 5944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 5431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508015783.750, "dur": 3038.565,
    "args": {
      "External id": 5945,"Record function id": 0, "Sequence number": 4699, "Fwd thread id": 1, "Ev Idx": 5432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508015788.222, "dur": 2860.313,
    "args": {
      "External id": 5946,"Sequence number": 4699, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["float", "", "", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 5433
    }
  },
  {
    "ph": "f", "id": 15, "pid": 14565, "tid": 14640, "ts": 1290508015788.222,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 14565, "tid": 14640,
    "ts": 1290508016065.045, "dur": 2547.369,
    "args": {
      "External id": 5947,"Record function id": 0, "scope": "user_scope", "Ev Idx": 5434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14640,
    "ts": 1290508017089.246, "dur": 36.353,
    "args": {
      "External id": 5948,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[]], [[]], []], "Input Dims": [[[]], [[]], []], "Ev Idx": 5435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018847.497, "dur": 18.842,
    "args": {
      "External id": 5949,"Record function id": 0, "Ev Idx": 5436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018855.332, "dur": 9.699,
    "args": {
      "External id": 5950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018859.777, "dur": 4.448,
    "args": {
      "External id": 5951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018861.858, "dur": 2.176,
    "args": {
      "External id": 5952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018871.115, "dur": 6.497,
    "args": {
      "External id": 5953,"Record function id": 0, "Ev Idx": 5440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018873.179, "dur": 3.608,
    "args": {
      "External id": 5954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018874.445, "dur": 1.998,
    "args": {
      "External id": 5955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018875.412, "dur": 0.873,
    "args": {
      "External id": 5956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018881.474, "dur": 5.250,
    "args": {
      "External id": 5957,"Record function id": 0, "Ev Idx": 5444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018883.148, "dur": 2.812,
    "args": {
      "External id": 5958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018884.052, "dur": 1.491,
    "args": {
      "External id": 5959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018884.621, "dur": 0.778,
    "args": {
      "External id": 5960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018890.086, "dur": 5.324,
    "args": {
      "External id": 5961,"Record function id": 0, "Ev Idx": 5448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018891.623, "dur": 3.019,
    "args": {
      "External id": 5962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018892.638, "dur": 1.522,
    "args": {
      "External id": 5963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018893.167, "dur": 0.823,
    "args": {
      "External id": 5964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018898.878, "dur": 4.542,
    "args": {
      "External id": 5965,"Record function id": 0, "Ev Idx": 5452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018900.276, "dur": 2.397,
    "args": {
      "External id": 5966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018901.070, "dur": 1.272,
    "args": {
      "External id": 5967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018901.433, "dur": 0.755,
    "args": {
      "External id": 5968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018906.693, "dur": 5.895,
    "args": {
      "External id": 5969,"Record function id": 0, "Ev Idx": 5456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018908.479, "dur": 3.362,
    "args": {
      "External id": 5970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018909.504, "dur": 2.055,
    "args": {
      "External id": 5971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508018910.130, "dur": 1.261,
    "args": {
      "External id": 5972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018916.166, "dur": 126.856,
    "args": {
      "External id": 5973,"Record function id": 0, "Ev Idx": 5460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508018917.693, "dur": 120.965,
    "args": {
      "External id": 5974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019027.721, "dur": 8.179,
    "args": {
      "External id": 5975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019031.016, "dur": 4.337,
    "args": {
      "External id": 5976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019064.748, "dur": 12.148,
    "args": {
      "External id": 5977,"Record function id": 0, "Ev Idx": 5464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019068.956, "dur": 7.048,
    "args": {
      "External id": 5978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019072.433, "dur": 2.943,
    "args": {
      "External id": 5979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019073.824, "dur": 1.400,
    "args": {
      "External id": 5980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019085.016, "dur": 6.022,
    "args": {
      "External id": 5981,"Record function id": 0, "Ev Idx": 5468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019086.965, "dur": 3.253,
    "args": {
      "External id": 5982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019088.009, "dur": 1.694,
    "args": {
      "External id": 5983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019088.637, "dur": 0.914,
    "args": {
      "External id": 5984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019095.000, "dur": 6.329,
    "args": {
      "External id": 5985,"Record function id": 0, "Ev Idx": 5472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019097.167, "dur": 3.427,
    "args": {
      "External id": 5986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019098.163, "dur": 1.078,
    "args": {
      "External id": 5987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019098.564, "dur": 0.550,
    "args": {
      "External id": 5988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019105.026, "dur": 5.220,
    "args": {
      "External id": 5989,"Record function id": 0, "Ev Idx": 5476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019106.849, "dur": 2.688,
    "args": {
      "External id": 5990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019107.753, "dur": 1.198,
    "args": {
      "External id": 5991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019108.211, "dur": 0.577,
    "args": {
      "External id": 5992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019113.948, "dur": 4.945,
    "args": {
      "External id": 5993,"Record function id": 0, "Ev Idx": 5480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019115.483, "dur": 2.537,
    "args": {
      "External id": 5994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019116.272, "dur": 1.214,
    "args": {
      "External id": 5995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019116.715, "dur": 0.647,
    "args": {
      "External id": 5996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019122.572, "dur": 4.702,
    "args": {
      "External id": 5997,"Record function id": 0, "Ev Idx": 5484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019124.058, "dur": 2.499,
    "args": {
      "External id": 5998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019124.770, "dur": 1.259,
    "args": {
      "External id": 5999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019125.156, "dur": 0.746,
    "args": {
      "External id": 6000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019130.956, "dur": 5.327,
    "args": {
      "External id": 6001,"Record function id": 0, "Ev Idx": 5488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019132.634, "dur": 2.830,
    "args": {
      "External id": 6002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019133.250, "dur": 1.560,
    "args": {
      "External id": 6003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019133.656, "dur": 1.022,
    "args": {
      "External id": 6004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019140.023, "dur": 6.526,
    "args": {
      "External id": 6005,"Record function id": 0, "Ev Idx": 5492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019143.141, "dur": 2.541,
    "args": {
      "External id": 6006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019143.766, "dur": 1.339,
    "args": {
      "External id": 6007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019144.149, "dur": 0.829,
    "args": {
      "External id": 6008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019150.135, "dur": 4.967,
    "args": {
      "External id": 6009,"Record function id": 0, "Ev Idx": 5496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019151.934, "dur": 2.485,
    "args": {
      "External id": 6010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019152.553, "dur": 1.338,
    "args": {
      "External id": 6011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019152.929, "dur": 0.830,
    "args": {
      "External id": 6012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019158.637, "dur": 4.769,
    "args": {
      "External id": 6013,"Record function id": 0, "Ev Idx": 5500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019160.143, "dur": 2.395,
    "args": {
      "External id": 6014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019160.871, "dur": 1.134,
    "args": {
      "External id": 6015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019161.281, "dur": 0.545,
    "args": {
      "External id": 6016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019167.289, "dur": 4.951,
    "args": {
      "External id": 6017,"Record function id": 0, "Ev Idx": 5504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019168.748, "dur": 2.651,
    "args": {
      "External id": 6018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019169.455, "dur": 1.404,
    "args": {
      "External id": 6019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019169.866, "dur": 0.848,
    "args": {
      "External id": 6020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019175.739, "dur": 4.452,
    "args": {
      "External id": 6021,"Record function id": 0, "Ev Idx": 5508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019177.015, "dur": 2.350,
    "args": {
      "External id": 6022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019177.618, "dur": 1.166,
    "args": {
      "External id": 6023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019178.109, "dur": 0.547,
    "args": {
      "External id": 6024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019183.876, "dur": 4.826,
    "args": {
      "External id": 6025,"Record function id": 0, "Ev Idx": 5512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019185.192, "dur": 2.807,
    "args": {
      "External id": 6026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019185.966, "dur": 1.486,
    "args": {
      "External id": 6027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019186.329, "dur": 0.986,
    "args": {
      "External id": 6028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019192.446, "dur": 5.237,
    "args": {
      "External id": 6029,"Record function id": 0, "Ev Idx": 5516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019194.474, "dur": 2.456,
    "args": {
      "External id": 6030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019195.262, "dur": 1.038,
    "args": {
      "External id": 6031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019195.647, "dur": 0.521,
    "args": {
      "External id": 6032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019201.337, "dur": 5.576,
    "args": {
      "External id": 6033,"Record function id": 0, "Ev Idx": 5520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019203.304, "dur": 2.764,
    "args": {
      "External id": 6034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019204.114, "dur": 1.335,
    "args": {
      "External id": 6035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019204.662, "dur": 0.649,
    "args": {
      "External id": 6036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019210.469, "dur": 4.687,
    "args": {
      "External id": 6037,"Record function id": 0, "Ev Idx": 5524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019212.014, "dur": 2.442,
    "args": {
      "External id": 6038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019212.626, "dur": 1.298,
    "args": {
      "External id": 6039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019213.026, "dur": 0.755,
    "args": {
      "External id": 6040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019218.903, "dur": 4.825,
    "args": {
      "External id": 6041,"Record function id": 0, "Ev Idx": 5528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019220.609, "dur": 2.426,
    "args": {
      "External id": 6042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019221.217, "dur": 1.281,
    "args": {
      "External id": 6043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019221.559, "dur": 0.811,
    "args": {
      "External id": 6044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019227.229, "dur": 5.012,
    "args": {
      "External id": 6045,"Record function id": 0, "Ev Idx": 5532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019228.907, "dur": 2.624,
    "args": {
      "External id": 6046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019229.555, "dur": 1.297,
    "args": {
      "External id": 6047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019229.931, "dur": 0.778,
    "args": {
      "External id": 6048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019236.089, "dur": 4.785,
    "args": {
      "External id": 6049,"Record function id": 0, "Ev Idx": 5536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019237.767, "dur": 2.417,
    "args": {
      "External id": 6050,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019238.641, "dur": 1.037,
    "args": {
      "External id": 6051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019238.974, "dur": 0.572,
    "args": {
      "External id": 6052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019244.684, "dur": 4.573,
    "args": {
      "External id": 6053,"Record function id": 0, "Ev Idx": 5540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019246.013, "dur": 2.538,
    "args": {
      "External id": 6054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019246.765, "dur": 1.136,
    "args": {
      "External id": 6055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019247.177, "dur": 0.577,
    "args": {
      "External id": 6056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019252.773, "dur": 4.839,
    "args": {
      "External id": 6057,"Record function id": 0, "Ev Idx": 5544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019254.255, "dur": 2.524,
    "args": {
      "External id": 6058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019254.862, "dur": 1.382,
    "args": {
      "External id": 6059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019255.242, "dur": 0.877,
    "args": {
      "External id": 6060,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019261.160, "dur": 4.779,
    "args": {
      "External id": 6061,"Record function id": 0, "Ev Idx": 5548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019262.862, "dur": 2.257,
    "args": {
      "External id": 6062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019263.470, "dur": 1.172,
    "args": {
      "External id": 6063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019263.823, "dur": 0.687,
    "args": {
      "External id": 6064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019269.844, "dur": 4.375,
    "args": {
      "External id": 6065,"Record function id": 0, "Ev Idx": 5552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019271.177, "dur": 2.327,
    "args": {
      "External id": 6066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019271.772, "dur": 1.166,
    "args": {
      "External id": 6067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019272.150, "dur": 0.660,
    "args": {
      "External id": 6068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019277.914, "dur": 4.526,
    "args": {
      "External id": 6069,"Record function id": 0, "Ev Idx": 5556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019279.236, "dur": 2.347,
    "args": {
      "External id": 6070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019279.990, "dur": 1.096,
    "args": {
      "External id": 6071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019280.340, "dur": 0.605,
    "args": {
      "External id": 6072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019286.190, "dur": 5.278,
    "args": {
      "External id": 6073,"Record function id": 0, "Ev Idx": 5560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019287.954, "dur": 2.788,
    "args": {
      "External id": 6074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019288.947, "dur": 1.222,
    "args": {
      "External id": 6075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019289.278, "dur": 0.758,
    "args": {
      "External id": 6076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019295.141, "dur": 5.382,
    "args": {
      "External id": 6077,"Record function id": 0, "Ev Idx": 5564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019297.214, "dur": 2.474,
    "args": {
      "External id": 6078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019297.921, "dur": 1.205,
    "args": {
      "External id": 6079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019298.271, "dur": 0.707,
    "args": {
      "External id": 6080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019303.794, "dur": 4.524,
    "args": {
      "External id": 6081,"Record function id": 0, "Ev Idx": 5568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019305.118, "dur": 2.515,
    "args": {
      "External id": 6082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019305.857, "dur": 1.277,
    "args": {
      "External id": 6083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019306.249, "dur": 0.744,
    "args": {
      "External id": 6084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019311.800, "dur": 5.232,
    "args": {
      "External id": 6085,"Record function id": 0, "Ev Idx": 5572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019313.272, "dur": 3.063,
    "args": {
      "External id": 6086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019314.129, "dur": 1.684,
    "args": {
      "External id": 6087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019314.666, "dur": 1.011,
    "args": {
      "External id": 6088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019320.921, "dur": 4.472,
    "args": {
      "External id": 6089,"Record function id": 0, "Ev Idx": 5576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019322.276, "dur": 2.263,
    "args": {
      "External id": 6090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019322.852, "dur": 1.088,
    "args": {
      "External id": 6091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019323.241, "dur": 0.573,
    "args": {
      "External id": 6092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019329.256, "dur": 5.315,
    "args": {
      "External id": 6093,"Record function id": 0, "Ev Idx": 5580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019330.887, "dur": 2.843,
    "args": {
      "External id": 6094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019331.714, "dur": 1.520,
    "args": {
      "External id": 6095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019332.054, "dur": 0.986,
    "args": {
      "External id": 6096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019338.200, "dur": 4.242,
    "args": {
      "External id": 6097,"Record function id": 0, "Ev Idx": 5584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019339.640, "dur": 2.100,
    "args": {
      "External id": 6098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019340.221, "dur": 1.026,
    "args": {
      "External id": 6099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019340.597, "dur": 0.503,
    "args": {
      "External id": 6100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019346.169, "dur": 4.699,
    "args": {
      "External id": 6101,"Record function id": 0, "Ev Idx": 5588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019347.504, "dur": 2.510,
    "args": {
      "External id": 6102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019348.466, "dur": 1.078,
    "args": {
      "External id": 6103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019348.836, "dur": 0.577,
    "args": {
      "External id": 6104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019354.421, "dur": 4.548,
    "args": {
      "External id": 6105,"Record function id": 0, "Ev Idx": 5592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019355.708, "dur": 2.420,
    "args": {
      "External id": 6106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019356.500, "dur": 1.099,
    "args": {
      "External id": 6107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019356.880, "dur": 0.580,
    "args": {
      "External id": 6108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019362.651, "dur": 4.930,
    "args": {
      "External id": 6109,"Record function id": 0, "Ev Idx": 5596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019363.924, "dur": 2.958,
    "args": {
      "External id": 6110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019364.655, "dur": 1.749,
    "args": {
      "External id": 6111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019365.037, "dur": 1.245,
    "args": {
      "External id": 6112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019371.279, "dur": 4.686,
    "args": {
      "External id": 6113,"Record function id": 0, "Ev Idx": 5600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019372.712, "dur": 2.407,
    "args": {
      "External id": 6114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019373.505, "dur": 1.135,
    "args": {
      "External id": 6115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019373.844, "dur": 0.663,
    "args": {
      "External id": 6116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019379.527, "dur": 4.917,
    "args": {
      "External id": 6117,"Record function id": 0, "Ev Idx": 5604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019381.386, "dur": 2.325,
    "args": {
      "External id": 6118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019382.205, "dur": 0.930,
    "args": {
      "External id": 6119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019382.566, "dur": 0.371,
    "args": {
      "External id": 6120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019387.978, "dur": 4.725,
    "args": {
      "External id": 6121,"Record function id": 0, "Ev Idx": 5608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019389.373, "dur": 2.464,
    "args": {
      "External id": 6122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019390.067, "dur": 1.168,
    "args": {
      "External id": 6123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019390.435, "dur": 0.670,
    "args": {
      "External id": 6124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019396.365, "dur": 4.884,
    "args": {
      "External id": 6125,"Record function id": 0, "Ev Idx": 5612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019397.798, "dur": 2.747,
    "args": {
      "External id": 6126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019398.457, "dur": 1.587,
    "args": {
      "External id": 6127,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019398.860, "dur": 1.047,
    "args": {
      "External id": 6128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019404.806, "dur": 4.735,
    "args": {
      "External id": 6129,"Record function id": 0, "Ev Idx": 5616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019406.176, "dur": 2.537,
    "args": {
      "External id": 6130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019406.874, "dur": 1.327,
    "args": {
      "External id": 6131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019407.211, "dur": 0.859,
    "args": {
      "External id": 6132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019413.265, "dur": 5.493,
    "args": {
      "External id": 6133,"Record function id": 0, "Ev Idx": 5620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019414.538, "dur": 3.362,
    "args": {
      "External id": 6134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019415.396, "dur": 1.936,
    "args": {
      "External id": 6135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019415.830, "dur": 1.362,
    "args": {
      "External id": 6136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019422.795, "dur": 4.526,
    "args": {
      "External id": 6137,"Record function id": 0, "Ev Idx": 5624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019424.234, "dur": 2.284,
    "args": {
      "External id": 6138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019424.979, "dur": 1.064,
    "args": {
      "External id": 6139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019425.368, "dur": 0.538,
    "args": {
      "External id": 6140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019430.821, "dur": 4.442,
    "args": {
      "External id": 6141,"Record function id": 0, "Ev Idx": 5628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019432.065, "dur": 2.386,
    "args": {
      "External id": 6142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019432.788, "dur": 1.151,
    "args": {
      "External id": 6143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019433.318, "dur": 0.484,
    "args": {
      "External id": 6144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019698.027, "dur": 15.294,
    "args": {
      "External id": 6145,"Record function id": 0, "Ev Idx": 5632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019707.069, "dur": 4.977,
    "args": {
      "External id": 6146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019708.697, "dur": 2.538,
    "args": {
      "External id": 6147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019709.537, "dur": 1.508,
    "args": {
      "External id": 6148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019721.653, "dur": 5.869,
    "args": {
      "External id": 6149,"Record function id": 0, "Ev Idx": 5636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019723.851, "dur": 2.730,
    "args": {
      "External id": 6150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019724.675, "dur": 1.405,
    "args": {
      "External id": 6151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019725.407, "dur": 0.525,
    "args": {
      "External id": 6152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019731.448, "dur": 5.338,
    "args": {
      "External id": 6153,"Record function id": 0, "Ev Idx": 5640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019733.360, "dur": 2.596,
    "args": {
      "External id": 6154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019734.015, "dur": 1.451,
    "args": {
      "External id": 6155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019734.518, "dur": 0.825,
    "args": {
      "External id": 6156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019740.925, "dur": 4.684,
    "args": {
      "External id": 6157,"Record function id": 0, "Ev Idx": 5644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019742.506, "dur": 2.272,
    "args": {
      "External id": 6158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019743.302, "dur": 1.018,
    "args": {
      "External id": 6159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019743.662, "dur": 0.536,
    "args": {
      "External id": 6160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019749.674, "dur": 5.310,
    "args": {
      "External id": 6161,"Record function id": 0, "Ev Idx": 5648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019751.804, "dur": 2.450,
    "args": {
      "External id": 6162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019752.539, "dur": 0.843,
    "args": {
      "External id": 6163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019752.899, "dur": 0.358,
    "args": {
      "External id": 6164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019758.929, "dur": 5.547,
    "args": {
      "External id": 6165,"Record function id": 0, "Ev Idx": 5652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019761.058, "dur": 2.575,
    "args": {
      "External id": 6166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019761.796, "dur": 1.261,
    "args": {
      "External id": 6167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019762.167, "dur": 0.764,
    "args": {
      "External id": 6168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019768.413, "dur": 4.736,
    "args": {
      "External id": 6169,"Record function id": 0, "Ev Idx": 5656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019770.120, "dur": 2.314,
    "args": {
      "External id": 6170,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019770.815, "dur": 1.149,
    "args": {
      "External id": 6171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019771.206, "dur": 0.637,
    "args": {
      "External id": 6172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019777.113, "dur": 5.196,
    "args": {
      "External id": 6173,"Record function id": 0, "Ev Idx": 5660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019778.953, "dur": 2.661,
    "args": {
      "External id": 6174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019779.633, "dur": 1.504,
    "args": {
      "External id": 6175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019779.994, "dur": 1.019,
    "args": {
      "External id": 6176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019785.974, "dur": 4.764,
    "args": {
      "External id": 6177,"Record function id": 0, "Ev Idx": 5664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019787.386, "dur": 2.499,
    "args": {
      "External id": 6178,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019788.033, "dur": 1.385,
    "args": {
      "External id": 6179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019788.440, "dur": 0.856,
    "args": {
      "External id": 6180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019794.543, "dur": 4.850,
    "args": {
      "External id": 6181,"Record function id": 0, "Ev Idx": 5668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019796.107, "dur": 2.574,
    "args": {
      "External id": 6182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019796.798, "dur": 1.408,
    "args": {
      "External id": 6183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019797.263, "dur": 0.812,
    "args": {
      "External id": 6184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019803.178, "dur": 4.531,
    "args": {
      "External id": 6185,"Record function id": 0, "Ev Idx": 5672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019804.656, "dur": 2.303,
    "args": {
      "External id": 6186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019805.252, "dur": 1.140,
    "args": {
      "External id": 6187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019805.652, "dur": 0.615,
    "args": {
      "External id": 6188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019811.445, "dur": 4.482,
    "args": {
      "External id": 6189,"Record function id": 0, "Ev Idx": 5676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019812.954, "dur": 2.273,
    "args": {
      "External id": 6190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019813.599, "dur": 1.124,
    "args": {
      "External id": 6191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019813.975, "dur": 0.627,
    "args": {
      "External id": 6192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019819.730, "dur": 4.998,
    "args": {
      "External id": 6193,"Record function id": 0, "Ev Idx": 5680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019821.418, "dur": 2.477,
    "args": {
      "External id": 6194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019822.169, "dur": 1.229,
    "args": {
      "External id": 6195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019822.758, "dur": 0.516,
    "args": {
      "External id": 6196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019828.420, "dur": 5.076,
    "args": {
      "External id": 6197,"Record function id": 0, "Ev Idx": 5684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019830.167, "dur": 2.625,
    "args": {
      "External id": 6198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019831.100, "dur": 1.227,
    "args": {
      "External id": 6199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019831.559, "dur": 0.642,
    "args": {
      "External id": 6200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019836.857, "dur": 4.530,
    "args": {
      "External id": 6201,"Record function id": 0, "Ev Idx": 5688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019838.409, "dur": 2.142,
    "args": {
      "External id": 6202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019839.007, "dur": 1.062,
    "args": {
      "External id": 6203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019839.373, "dur": 0.571,
    "args": {
      "External id": 6204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019845.153, "dur": 5.187,
    "args": {
      "External id": 6205,"Record function id": 0, "Ev Idx": 5692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019847.267, "dur": 2.356,
    "args": {
      "External id": 6206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019847.985, "dur": 0.979,
    "args": {
      "External id": 6207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019848.351, "dur": 0.490,
    "args": {
      "External id": 6208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019853.787, "dur": 4.746,
    "args": {
      "External id": 6209,"Record function id": 0, "Ev Idx": 5696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019855.227, "dur": 2.442,
    "args": {
      "External id": 6210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019856.120, "dur": 0.953,
    "args": {
      "External id": 6211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019856.487, "dur": 0.459,
    "args": {
      "External id": 6212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019861.757, "dur": 4.527,
    "args": {
      "External id": 6213,"Record function id": 0, "Ev Idx": 5700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019863.134, "dur": 2.303,
    "args": {
      "External id": 6214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019863.791, "dur": 1.137,
    "args": {
      "External id": 6215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019864.189, "dur": 0.614,
    "args": {
      "External id": 6216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019869.989, "dur": 4.774,
    "args": {
      "External id": 6217,"Record function id": 0, "Ev Idx": 5704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019871.646, "dur": 2.262,
    "args": {
      "External id": 6218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019872.323, "dur": 1.108,
    "args": {
      "External id": 6219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019872.722, "dur": 0.567,
    "args": {
      "External id": 6220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019878.380, "dur": 4.587,
    "args": {
      "External id": 6221,"Record function id": 0, "Ev Idx": 5708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019879.761, "dur": 2.349,
    "args": {
      "External id": 6222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019880.456, "dur": 1.037,
    "args": {
      "External id": 6223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019880.850, "dur": 0.503,
    "args": {
      "External id": 6224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019886.536, "dur": 4.518,
    "args": {
      "External id": 6225,"Record function id": 0, "Ev Idx": 5712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019888.037, "dur": 2.185,
    "args": {
      "External id": 6226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019888.697, "dur": 1.037,
    "args": {
      "External id": 6227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019889.059, "dur": 0.554,
    "args": {
      "External id": 6228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019894.712, "dur": 4.614,
    "args": {
      "External id": 6229,"Record function id": 0, "Ev Idx": 5716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019896.124, "dur": 2.326,
    "args": {
      "External id": 6230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019896.842, "dur": 1.077,
    "args": {
      "External id": 6231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019897.214, "dur": 0.582,
    "args": {
      "External id": 6232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019902.844, "dur": 9.918,
    "args": {
      "External id": 6233,"Record function id": 0, "Ev Idx": 5720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019904.597, "dur": 7.311,
    "args": {
      "External id": 6234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019909.939, "dur": 1.401,
    "args": {
      "External id": 6235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019910.325, "dur": 0.884,
    "args": {
      "External id": 6236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019916.577, "dur": 5.332,
    "args": {
      "External id": 6237,"Record function id": 0, "Ev Idx": 5724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019918.523, "dur": 2.683,
    "args": {
      "External id": 6238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019919.524, "dur": 1.029,
    "args": {
      "External id": 6239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019919.909, "dur": 0.522,
    "args": {
      "External id": 6240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019925.569, "dur": 5.236,
    "args": {
      "External id": 6241,"Record function id": 0, "Ev Idx": 5728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019927.596, "dur": 2.355,
    "args": {
      "External id": 6242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019928.354, "dur": 0.964,
    "args": {
      "External id": 6243,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019928.712, "dur": 0.485,
    "args": {
      "External id": 6244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019934.422, "dur": 4.483,
    "args": {
      "External id": 6245,"Record function id": 0, "Ev Idx": 5732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019935.867, "dur": 2.193,
    "args": {
      "External id": 6246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019936.427, "dur": 1.133,
    "args": {
      "External id": 6247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019937.049, "dur": 0.358,
    "args": {
      "External id": 6248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019942.519, "dur": 4.798,
    "args": {
      "External id": 6249,"Record function id": 0, "Ev Idx": 5736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019944.350, "dur": 2.249,
    "args": {
      "External id": 6250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019945.133, "dur": 0.864,
    "args": {
      "External id": 6251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019945.507, "dur": 0.352,
    "args": {
      "External id": 6252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019950.774, "dur": 4.377,
    "args": {
      "External id": 6253,"Record function id": 0, "Ev Idx": 5740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019952.281, "dur": 2.020,
    "args": {
      "External id": 6254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019952.951, "dur": 0.840,
    "args": {
      "External id": 6255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019953.346, "dur": 0.318,
    "args": {
      "External id": 6256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019958.748, "dur": 4.400,
    "args": {
      "External id": 6257,"Record function id": 0, "Ev Idx": 5744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019960.148, "dur": 2.259,
    "args": {
      "External id": 6258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019960.783, "dur": 1.013,
    "args": {
      "External id": 6259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019961.163, "dur": 0.507,
    "args": {
      "External id": 6260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019966.687, "dur": 4.855,
    "args": {
      "External id": 6261,"Record function id": 0, "Ev Idx": 5748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019968.165, "dur": 2.669,
    "args": {
      "External id": 6262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019968.853, "dur": 1.398,
    "args": {
      "External id": 6263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019969.258, "dur": 0.870,
    "args": {
      "External id": 6264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019975.256, "dur": 5.036,
    "args": {
      "External id": 6265,"Record function id": 0, "Ev Idx": 5752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019976.786, "dur": 2.649,
    "args": {
      "External id": 6266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019977.464, "dur": 1.428,
    "args": {
      "External id": 6267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019977.831, "dur": 0.925,
    "args": {
      "External id": 6268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019983.566, "dur": 4.499,
    "args": {
      "External id": 6269,"Record function id": 0, "Ev Idx": 5756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019984.934, "dur": 2.426,
    "args": {
      "External id": 6270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019985.516, "dur": 1.321,
    "args": {
      "External id": 6271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019985.888, "dur": 0.786,
    "args": {
      "External id": 6272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019991.582, "dur": 4.606,
    "args": {
      "External id": 6273,"Record function id": 0, "Ev Idx": 5760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019993.028, "dur": 2.314,
    "args": {
      "External id": 6274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019993.755, "dur": 0.960,
    "args": {
      "External id": 6275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508019994.137, "dur": 0.457,
    "args": {
      "External id": 6276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508019999.594, "dur": 4.170,
    "args": {
      "External id": 6277,"Record function id": 0, "Ev Idx": 5764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020000.921, "dur": 2.138,
    "args": {
      "External id": 6278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020001.474, "dur": 1.092,
    "args": {
      "External id": 6279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020001.849, "dur": 0.592,
    "args": {
      "External id": 6280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020007.514, "dur": 4.424,
    "args": {
      "External id": 6281,"Record function id": 0, "Ev Idx": 5768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020008.839, "dur": 2.264,
    "args": {
      "External id": 6282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020009.491, "dur": 1.094,
    "args": {
      "External id": 6283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020009.864, "dur": 0.597,
    "args": {
      "External id": 6284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020015.603, "dur": 4.344,
    "args": {
      "External id": 6285,"Record function id": 0, "Ev Idx": 5772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020016.918, "dur": 2.290,
    "args": {
      "External id": 6286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020017.599, "dur": 1.131,
    "args": {
      "External id": 6287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020017.967, "dur": 0.641,
    "args": {
      "External id": 6288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020023.549, "dur": 5.502,
    "args": {
      "External id": 6289,"Record function id": 0, "Ev Idx": 5776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020025.399, "dur": 2.885,
    "args": {
      "External id": 6290,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020026.164, "dur": 1.536,
    "args": {
      "External id": 6291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020026.529, "dur": 1.046,
    "args": {
      "External id": 6292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020032.491, "dur": 4.986,
    "args": {
      "External id": 6293,"Record function id": 0, "Ev Idx": 5780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020034.292, "dur": 2.436,
    "args": {
      "External id": 6294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020034.966, "dur": 1.049,
    "args": {
      "External id": 6295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020035.432, "dur": 0.450,
    "args": {
      "External id": 6296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020041.248, "dur": 4.017,
    "args": {
      "External id": 6297,"Record function id": 0, "Ev Idx": 5784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020042.570, "dur": 1.990,
    "args": {
      "External id": 6298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020043.122, "dur": 0.943,
    "args": {
      "External id": 6299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020043.495, "dur": 0.445,
    "args": {
      "External id": 6300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020048.689, "dur": 7.780,
    "args": {
      "External id": 6301,"Record function id": 0, "Ev Idx": 5788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020049.965, "dur": 5.671,
    "args": {
      "External id": 6302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020050.756, "dur": 4.365,
    "args": {
      "External id": 6303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020054.672, "dur": 0.312,
    "args": {
      "External id": 6304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020060.196, "dur": 4.272,
    "args": {
      "External id": 6305,"Record function id": 0, "Ev Idx": 5792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020061.606, "dur": 2.017,
    "args": {
      "External id": 6306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020062.232, "dur": 0.861,
    "args": {
      "External id": 6307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020062.624, "dur": 0.349,
    "args": {
      "External id": 6308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020068.220, "dur": 4.268,
    "args": {
      "External id": 6309,"Record function id": 0, "Ev Idx": 5796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020069.795, "dur": 2.000,
    "args": {
      "External id": 6310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020070.411, "dur": 0.831,
    "args": {
      "External id": 6311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020070.768, "dur": 0.350,
    "args": {
      "External id": 6312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020076.027, "dur": 7.952,
    "args": {
      "External id": 6313,"Record function id": 0, "Ev Idx": 5800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020077.657, "dur": 2.167,
    "args": {
      "External id": 6314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020078.413, "dur": 0.914,
    "args": {
      "External id": 6315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020078.787, "dur": 0.404,
    "args": {
      "External id": 6316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020087.582, "dur": 4.304,
    "args": {
      "External id": 6317,"Record function id": 0, "Ev Idx": 5804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020088.975, "dur": 2.068,
    "args": {
      "External id": 6318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020089.567, "dur": 0.942,
    "args": {
      "External id": 6319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020089.966, "dur": 0.407,
    "args": {
      "External id": 6320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020095.450, "dur": 4.625,
    "args": {
      "External id": 6321,"Record function id": 0, "Ev Idx": 5808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020097.039, "dur": 2.326,
    "args": {
      "External id": 6322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020097.698, "dur": 1.109,
    "args": {
      "External id": 6323,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020098.097, "dur": 0.571,
    "args": {
      "External id": 6324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020103.796, "dur": 4.526,
    "args": {
      "External id": 6325,"Record function id": 0, "Ev Idx": 5812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020105.282, "dur": 2.192,
    "args": {
      "External id": 6326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020106.025, "dur": 0.961,
    "args": {
      "External id": 6327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020106.506, "dur": 0.347,
    "args": {
      "External id": 6328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020112.080, "dur": 5.093,
    "args": {
      "External id": 6329,"Record function id": 0, "Ev Idx": 5816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020113.714, "dur": 2.755,
    "args": {
      "External id": 6330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020114.563, "dur": 1.383,
    "args": {
      "External id": 6331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020115.064, "dur": 0.750,
    "args": {
      "External id": 6332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 5819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020120.994, "dur": 4.773,
    "args": {
      "External id": 6333,"Record function id": 0, "Ev Idx": 5820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020122.634, "dur": 2.430,
    "args": {
      "External id": 6334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020123.358, "dur": 1.138,
    "args": {
      "External id": 6335,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020123.809, "dur": 0.524,
    "args": {
      "External id": 6336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020129.429, "dur": 5.518,
    "args": {
      "External id": 6337,"Record function id": 0, "Ev Idx": 5824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020131.515, "dur": 2.731,
    "args": {
      "External id": 6338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020132.217, "dur": 1.517,
    "args": {
      "External id": 6339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020132.667, "dur": 0.940,
    "args": {
      "External id": 6340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020138.994, "dur": 5.062,
    "args": {
      "External id": 6341,"Record function id": 0, "Ev Idx": 5828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020140.519, "dur": 2.702,
    "args": {
      "External id": 6342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020141.180, "dur": 1.431,
    "args": {
      "External id": 6343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020141.579, "dur": 0.888,
    "args": {
      "External id": 6344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020147.479, "dur": 5.099,
    "args": {
      "External id": 6345,"Record function id": 0, "Ev Idx": 5832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020148.867, "dur": 3.005,
    "args": {
      "External id": 6346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020149.623, "dur": 1.700,
    "args": {
      "External id": 6347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020149.998, "dur": 1.193,
    "args": {
      "External id": 6348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020156.087, "dur": 4.253,
    "args": {
      "External id": 6349,"Record function id": 0, "Ev Idx": 5836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020157.369, "dur": 2.286,
    "args": {
      "External id": 6350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020158.069, "dur": 1.017,
    "args": {
      "External id": 6351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020158.446, "dur": 0.511,
    "args": {
      "External id": 6352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 5839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020163.805, "dur": 4.686,
    "args": {
      "External id": 6353,"Record function id": 0, "Ev Idx": 5840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020165.367, "dur": 2.410,
    "args": {
      "External id": 6354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 5841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020166.089, "dur": 1.144,
    "args": {
      "External id": 6355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 5842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020166.687, "dur": 0.399,
    "args": {
      "External id": 6356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33644, 32]], "Ev Idx": 5843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020172.167, "dur": 5.224,
    "args": {
      "External id": 6357,"Record function id": 0, "Ev Idx": 5844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020173.891, "dur": 2.657,
    "args": {
      "External id": 6358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020174.875, "dur": 1.165,
    "args": {
      "External id": 6359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020175.356, "dur": 0.519,
    "args": {
      "External id": 6360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020180.773, "dur": 5.383,
    "args": {
      "External id": 6361,"Record function id": 0, "Ev Idx": 5848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020182.708, "dur": 2.728,
    "args": {
      "External id": 6362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020183.595, "dur": 1.261,
    "args": {
      "External id": 6363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020184.060, "dur": 0.664,
    "args": {
      "External id": 6364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020189.356, "dur": 5.310,
    "args": {
      "External id": 6365,"Record function id": 0, "Ev Idx": 5852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020191.220, "dur": 2.579,
    "args": {
      "External id": 6366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020192.248, "dur": 0.953,
    "args": {
      "External id": 6367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020192.713, "dur": 0.357,
    "args": {
      "External id": 6368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020198.124, "dur": 5.012,
    "args": {
      "External id": 6369,"Record function id": 0, "Ev Idx": 5856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020199.942, "dur": 2.487,
    "args": {
      "External id": 6370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020200.760, "dur": 1.142,
    "args": {
      "External id": 6371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020201.164, "dur": 0.617,
    "args": {
      "External id": 6372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020206.574, "dur": 4.459,
    "args": {
      "External id": 6373,"Record function id": 0, "Ev Idx": 5860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020207.952, "dur": 2.258,
    "args": {
      "External id": 6374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020208.711, "dur": 0.979,
    "args": {
      "External id": 6375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020209.075, "dur": 0.483,
    "args": {
      "External id": 6376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020214.765, "dur": 10.496,
    "args": {
      "External id": 6377,"Record function id": 0, "Ev Idx": 5864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020216.327, "dur": 8.105,
    "args": {
      "External id": 6378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020216.979, "dur": 0.860,
    "args": {
      "External id": 6379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020217.352, "dur": 0.361,
    "args": {
      "External id": 6380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020228.753, "dur": 4.374,
    "args": {
      "External id": 6381,"Record function id": 0, "Ev Idx": 5868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020230.330, "dur": 2.117,
    "args": {
      "External id": 6382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020231.082, "dur": 0.828,
    "args": {
      "External id": 6383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020231.443, "dur": 0.345,
    "args": {
      "External id": 6384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020236.583, "dur": 4.655,
    "args": {
      "External id": 6385,"Record function id": 0, "Ev Idx": 5872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020238.170, "dur": 2.210,
    "args": {
      "External id": 6386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020238.812, "dur": 0.859,
    "args": {
      "External id": 6387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020239.199, "dur": 0.353,
    "args": {
      "External id": 6388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020244.744, "dur": 4.315,
    "args": {
      "External id": 6389,"Record function id": 0, "Ev Idx": 5876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020246.079, "dur": 2.309,
    "args": {
      "External id": 6390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020246.719, "dur": 1.159,
    "args": {
      "External id": 6391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020247.105, "dur": 0.650,
    "args": {
      "External id": 6392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020252.811, "dur": 4.620,
    "args": {
      "External id": 6393,"Record function id": 0, "Ev Idx": 5880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020254.359, "dur": 2.354,
    "args": {
      "External id": 6394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020255.020, "dur": 1.139,
    "args": {
      "External id": 6395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020255.445, "dur": 0.591,
    "args": {
      "External id": 6396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020261.250, "dur": 4.817,
    "args": {
      "External id": 6397,"Record function id": 0, "Ev Idx": 5884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020262.925, "dur": 2.402,
    "args": {
      "External id": 6398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020263.678, "dur": 1.314,
    "args": {
      "External id": 6399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020264.056, "dur": 0.798,
    "args": {
      "External id": 6400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020269.495, "dur": 4.642,
    "args": {
      "External id": 6401,"Record function id": 0, "Ev Idx": 5888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020270.833, "dur": 2.468,
    "args": {
      "External id": 6402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020271.506, "dur": 1.269,
    "args": {
      "External id": 6403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020271.872, "dur": 0.777,
    "args": {
      "External id": 6404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020277.987, "dur": 5.374,
    "args": {
      "External id": 6405,"Record function id": 0, "Ev Idx": 5892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020279.906, "dur": 2.580,
    "args": {
      "External id": 6406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020280.781, "dur": 1.125,
    "args": {
      "External id": 6407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020281.148, "dur": 0.615,
    "args": {
      "External id": 6408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020286.971, "dur": 5.401,
    "args": {
      "External id": 6409,"Record function id": 0, "Ev Idx": 5896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020288.829, "dur": 2.824,
    "args": {
      "External id": 6410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020289.639, "dur": 1.575,
    "args": {
      "External id": 6411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020290.258, "dur": 0.820,
    "args": {
      "External id": 6412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020295.930, "dur": 5.085,
    "args": {
      "External id": 6413,"Record function id": 0, "Ev Idx": 5900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020297.353, "dur": 2.809,
    "args": {
      "External id": 6414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020298.076, "dur": 1.583,
    "args": {
      "External id": 6415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020298.451, "dur": 1.047,
    "args": {
      "External id": 6416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020304.885, "dur": 4.376,
    "args": {
      "External id": 6417,"Record function id": 0, "Ev Idx": 5904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020306.339, "dur": 2.218,
    "args": {
      "External id": 6418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020307.003, "dur": 1.014,
    "args": {
      "External id": 6419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020307.372, "dur": 0.511,
    "args": {
      "External id": 6420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020312.935, "dur": 4.866,
    "args": {
      "External id": 6421,"Record function id": 0, "Ev Idx": 5908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020314.627, "dur": 2.460,
    "args": {
      "External id": 6422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020315.363, "dur": 1.090,
    "args": {
      "External id": 6423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020315.741, "dur": 0.571,
    "args": {
      "External id": 6424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020321.613, "dur": 5.004,
    "args": {
      "External id": 6425,"Record function id": 0, "Ev Idx": 5912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020323.248, "dur": 2.661,
    "args": {
      "External id": 6426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020323.885, "dur": 1.482,
    "args": {
      "External id": 6427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020324.245, "dur": 0.995,
    "args": {
      "External id": 6428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020330.346, "dur": 4.580,
    "args": {
      "External id": 6429,"Record function id": 0, "Ev Idx": 5916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020332.016, "dur": 2.212,
    "args": {
      "External id": 6430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020332.691, "dur": 0.995,
    "args": {
      "External id": 6431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020333.065, "dur": 0.475,
    "args": {
      "External id": 6432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020338.411, "dur": 4.887,
    "args": {
      "External id": 6433,"Record function id": 0, "Ev Idx": 5920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020340.072, "dur": 2.394,
    "args": {
      "External id": 6434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020340.920, "dur": 1.005,
    "args": {
      "External id": 6435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020341.390, "dur": 0.396,
    "args": {
      "External id": 6436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020346.965, "dur": 4.447,
    "args": {
      "External id": 6437,"Record function id": 0, "Ev Idx": 5924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020348.404, "dur": 2.165,
    "args": {
      "External id": 6438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020349.092, "dur": 0.985,
    "args": {
      "External id": 6439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020349.452, "dur": 0.495,
    "args": {
      "External id": 6440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020355.060, "dur": 4.506,
    "args": {
      "External id": 6441,"Record function id": 0, "Ev Idx": 5928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020356.664, "dur": 2.217,
    "args": {
      "External id": 6442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020357.346, "dur": 1.036,
    "args": {
      "External id": 6443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020357.720, "dur": 0.541,
    "args": {
      "External id": 6444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020362.826, "dur": 4.630,
    "args": {
      "External id": 6445,"Record function id": 0, "Ev Idx": 5932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020364.470, "dur": 2.283,
    "args": {
      "External id": 6446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020365.202, "dur": 1.039,
    "args": {
      "External id": 6447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020365.569, "dur": 0.544,
    "args": {
      "External id": 6448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020370.959, "dur": 5.146,
    "args": {
      "External id": 6449,"Record function id": 0, "Ev Idx": 5936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020372.948, "dur": 2.437,
    "args": {
      "External id": 6450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020373.812, "dur": 1.007,
    "args": {
      "External id": 6451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020374.180, "dur": 0.519,
    "args": {
      "External id": 6452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020379.304, "dur": 4.677,
    "args": {
      "External id": 6453,"Record function id": 0, "Ev Idx": 5940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020381.119, "dur": 2.168,
    "args": {
      "External id": 6454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020381.843, "dur": 1.034,
    "args": {
      "External id": 6455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020382.266, "dur": 0.481,
    "args": {
      "External id": 6456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020387.449, "dur": 4.148,
    "args": {
      "External id": 6457,"Record function id": 0, "Ev Idx": 5944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020388.766, "dur": 2.133,
    "args": {
      "External id": 6458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020389.409, "dur": 0.954,
    "args": {
      "External id": 6459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020389.777, "dur": 0.464,
    "args": {
      "External id": 6460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020395.168, "dur": 5.473,
    "args": {
      "External id": 6461,"Record function id": 0, "Ev Idx": 5948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020397.421, "dur": 2.409,
    "args": {
      "External id": 6462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020398.099, "dur": 1.171,
    "args": {
      "External id": 6463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020398.491, "dur": 0.649,
    "args": {
      "External id": 6464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020403.996, "dur": 4.343,
    "args": {
      "External id": 6465,"Record function id": 0, "Ev Idx": 5952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020405.291, "dur": 2.331,
    "args": {
      "External id": 6466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020405.963, "dur": 1.054,
    "args": {
      "External id": 6467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020406.349, "dur": 0.542,
    "args": {
      "External id": 6468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020411.546, "dur": 4.226,
    "args": {
      "External id": 6469,"Record function id": 0, "Ev Idx": 5956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020412.825, "dur": 2.255,
    "args": {
      "External id": 6470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020413.516, "dur": 1.055,
    "args": {
      "External id": 6471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020413.879, "dur": 0.555,
    "args": {
      "External id": 6472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020419.373, "dur": 5.145,
    "args": {
      "External id": 6473,"Record function id": 0, "Ev Idx": 5960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020420.947, "dur": 2.861,
    "args": {
      "External id": 6474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020421.663, "dur": 1.657,
    "args": {
      "External id": 6475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020422.053, "dur": 1.125,
    "args": {
      "External id": 6476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020427.892, "dur": 5.410,
    "args": {
      "External id": 6477,"Record function id": 0, "Ev Idx": 5964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020429.493, "dur": 3.122,
    "args": {
      "External id": 6478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020430.212, "dur": 1.906,
    "args": {
      "External id": 6479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020430.594, "dur": 1.401,
    "args": {
      "External id": 6480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020436.929, "dur": 4.545,
    "args": {
      "External id": 6481,"Record function id": 0, "Ev Idx": 5968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020438.531, "dur": 2.235,
    "args": {
      "External id": 6482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020439.164, "dur": 1.057,
    "args": {
      "External id": 6483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020439.539, "dur": 0.557,
    "args": {
      "External id": 6484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020445.214, "dur": 4.329,
    "args": {
      "External id": 6485,"Record function id": 0, "Ev Idx": 5972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020446.571, "dur": 2.262,
    "args": {
      "External id": 6486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020447.215, "dur": 1.094,
    "args": {
      "External id": 6487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020447.613, "dur": 0.569,
    "args": {
      "External id": 6488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 5975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020453.065, "dur": 5.172,
    "args": {
      "External id": 6489,"Record function id": 0, "Ev Idx": 5976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020455.367, "dur": 2.153,
    "args": {
      "External id": 6490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020456.049, "dur": 0.993,
    "args": {
      "External id": 6491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020456.426, "dur": 0.486,
    "args": {
      "External id": 6492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 5979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020462.376, "dur": 5.231,
    "args": {
      "External id": 6493,"Record function id": 0, "Ev Idx": 5980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020464.166, "dur": 2.603,
    "args": {
      "External id": 6494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020465.037, "dur": 1.218,
    "args": {
      "External id": 6495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020465.499, "dur": 0.620,
    "args": {
      "External id": 6496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 5983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020471.550, "dur": 5.313,
    "args": {
      "External id": 6497,"Record function id": 0, "Ev Idx": 5984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020473.338, "dur": 2.670,
    "args": {
      "External id": 6498,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020474.063, "dur": 1.345,
    "args": {
      "External id": 6499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020474.511, "dur": 0.769,
    "args": {
      "External id": 6500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 5987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020480.713, "dur": 4.797,
    "args": {
      "External id": 6501,"Record function id": 0, "Ev Idx": 5988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020482.233, "dur": 2.437,
    "args": {
      "External id": 6502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020483.252, "dur": 1.117,
    "args": {
      "External id": 6503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020483.710, "dur": 0.531,
    "args": {
      "External id": 6504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 5991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020489.182, "dur": 4.774,
    "args": {
      "External id": 6505,"Record function id": 0, "Ev Idx": 5992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020490.656, "dur": 2.438,
    "args": {
      "External id": 6506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020491.287, "dur": 1.149,
    "args": {
      "External id": 6507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020491.664, "dur": 0.640,
    "args": {
      "External id": 6508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 5995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020497.805, "dur": 4.769,
    "args": {
      "External id": 6509,"Record function id": 0, "Ev Idx": 5996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020499.197, "dur": 2.525,
    "args": {
      "External id": 6510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020500.174, "dur": 1.026,
    "args": {
      "External id": 6511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020500.584, "dur": 0.484,
    "args": {
      "External id": 6512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 5999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020506.013, "dur": 4.562,
    "args": {
      "External id": 6513,"Record function id": 0, "Ev Idx": 6000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020507.668, "dur": 2.227,
    "args": {
      "External id": 6514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020508.368, "dur": 1.202,
    "args": {
      "External id": 6515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020508.778, "dur": 0.654,
    "args": {
      "External id": 6516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020514.119, "dur": 5.089,
    "args": {
      "External id": 6517,"Record function id": 0, "Ev Idx": 6004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020515.807, "dur": 2.559,
    "args": {
      "External id": 6518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020516.707, "dur": 1.024,
    "args": {
      "External id": 6519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020517.097, "dur": 0.510,
    "args": {
      "External id": 6520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020522.756, "dur": 4.849,
    "args": {
      "External id": 6521,"Record function id": 0, "Ev Idx": 6008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020524.389, "dur": 2.532,
    "args": {
      "External id": 6522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020525.271, "dur": 1.127,
    "args": {
      "External id": 6523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020525.639, "dur": 0.627,
    "args": {
      "External id": 6524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020531.171, "dur": 4.445,
    "args": {
      "External id": 6525,"Record function id": 0, "Ev Idx": 6012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020532.582, "dur": 2.345,
    "args": {
      "External id": 6526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020533.371, "dur": 1.261,
    "args": {
      "External id": 6527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020533.760, "dur": 0.741,
    "args": {
      "External id": 6528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020539.372, "dur": 5.031,
    "args": {
      "External id": 6529,"Record function id": 0, "Ev Idx": 6016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020541.092, "dur": 2.437,
    "args": {
      "External id": 6530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020541.774, "dur": 1.243,
    "args": {
      "External id": 6531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020542.140, "dur": 0.738,
    "args": {
      "External id": 6532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020548.211, "dur": 5.411,
    "args": {
      "External id": 6533,"Record function id": 0, "Ev Idx": 6020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020550.143, "dur": 2.733,
    "args": {
      "External id": 6534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020550.880, "dur": 1.341,
    "args": {
      "External id": 6535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020551.248, "dur": 0.838,
    "args": {
      "External id": 6536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020557.165, "dur": 5.717,
    "args": {
      "External id": 6537,"Record function id": 0, "Ev Idx": 6024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020559.255, "dur": 2.762,
    "args": {
      "External id": 6538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020560.115, "dur": 1.312,
    "args": {
      "External id": 6539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020560.485, "dur": 0.813,
    "args": {
      "External id": 6540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020566.339, "dur": 4.903,
    "args": {
      "External id": 6541,"Record function id": 0, "Ev Idx": 6028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020568.168, "dur": 2.256,
    "args": {
      "External id": 6542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020568.807, "dur": 1.116,
    "args": {
      "External id": 6543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020569.208, "dur": 0.584,
    "args": {
      "External id": 6544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020574.975, "dur": 5.024,
    "args": {
      "External id": 6545,"Record function id": 0, "Ev Idx": 6032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020576.416, "dur": 2.745,
    "args": {
      "External id": 6546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020577.150, "dur": 1.486,
    "args": {
      "External id": 6547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020577.519, "dur": 0.987,
    "args": {
      "External id": 6548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020583.883, "dur": 4.654,
    "args": {
      "External id": 6549,"Record function id": 0, "Ev Idx": 6036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020585.625, "dur": 2.202,
    "args": {
      "External id": 6550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020586.369, "dur": 0.959,
    "args": {
      "External id": 6551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020586.783, "dur": 0.396,
    "args": {
      "External id": 6552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020592.262, "dur": 5.147,
    "args": {
      "External id": 6553,"Record function id": 0, "Ev Idx": 6040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020593.990, "dur": 2.589,
    "args": {
      "External id": 6554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020594.768, "dur": 1.287,
    "args": {
      "External id": 6555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020595.138, "dur": 0.730,
    "args": {
      "External id": 6556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020601.080, "dur": 4.507,
    "args": {
      "External id": 6557,"Record function id": 0, "Ev Idx": 6044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020602.554, "dur": 2.352,
    "args": {
      "External id": 6558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020603.253, "dur": 1.050,
    "args": {
      "External id": 6559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020603.633, "dur": 0.525,
    "args": {
      "External id": 6560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020609.215, "dur": 4.197,
    "args": {
      "External id": 6561,"Record function id": 0, "Ev Idx": 6048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020610.513, "dur": 2.235,
    "args": {
      "External id": 6562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020611.175, "dur": 1.042,
    "args": {
      "External id": 6563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020611.558, "dur": 0.531,
    "args": {
      "External id": 6564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020617.125, "dur": 4.149,
    "args": {
      "External id": 6565,"Record function id": 0, "Ev Idx": 6052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020618.415, "dur": 2.168,
    "args": {
      "External id": 6566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020619.129, "dur": 0.962,
    "args": {
      "External id": 6567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020619.491, "dur": 0.413,
    "args": {
      "External id": 6568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020624.731, "dur": 4.893,
    "args": {
      "External id": 6569,"Record function id": 0, "Ev Idx": 6056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020626.009, "dur": 2.940,
    "args": {
      "External id": 6570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020626.803, "dur": 1.628,
    "args": {
      "External id": 6571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020627.383, "dur": 0.921,
    "args": {
      "External id": 6572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020633.389, "dur": 4.790,
    "args": {
      "External id": 6573,"Record function id": 0, "Ev Idx": 6060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020634.966, "dur": 2.530,
    "args": {
      "External id": 6574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020635.801, "dur": 1.185,
    "args": {
      "External id": 6575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020636.218, "dur": 0.583,
    "args": {
      "External id": 6576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020641.789, "dur": 5.569,
    "args": {
      "External id": 6577,"Record function id": 0, "Ev Idx": 6064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020643.794, "dur": 2.721,
    "args": {
      "External id": 6578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020644.884, "dur": 1.004,
    "args": {
      "External id": 6579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020645.406, "dur": 0.349,
    "args": {
      "External id": 6580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020650.852, "dur": 6.099,
    "args": {
      "External id": 6581,"Record function id": 0, "Ev Idx": 6068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020653.386, "dur": 2.858,
    "args": {
      "External id": 6582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020654.072, "dur": 1.504,
    "args": {
      "External id": 6583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020654.524, "dur": 0.868,
    "args": {
      "External id": 6584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020660.324, "dur": 4.416,
    "args": {
      "External id": 6585,"Record function id": 0, "Ev Idx": 6072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020661.618, "dur": 2.284,
    "args": {
      "External id": 6586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020662.207, "dur": 1.151,
    "args": {
      "External id": 6587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020662.693, "dur": 0.542,
    "args": {
      "External id": 6588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020667.939, "dur": 5.356,
    "args": {
      "External id": 6589,"Record function id": 0, "Ev Idx": 6076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020669.643, "dur": 2.973,
    "args": {
      "External id": 6590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020670.489, "dur": 1.607,
    "args": {
      "External id": 6591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020670.853, "dur": 1.066,
    "args": {
      "External id": 6592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020676.367, "dur": 4.737,
    "args": {
      "External id": 6593,"Record function id": 0, "Ev Idx": 6080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020677.923, "dur": 2.374,
    "args": {
      "External id": 6594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020678.614, "dur": 1.070,
    "args": {
      "External id": 6595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020678.998, "dur": 0.564,
    "args": {
      "External id": 6596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020684.379, "dur": 4.270,
    "args": {
      "External id": 6597,"Record function id": 0, "Ev Idx": 6084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020685.773, "dur": 2.178,
    "args": {
      "External id": 6598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020686.460, "dur": 0.939,
    "args": {
      "External id": 6599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020686.824, "dur": 0.403,
    "args": {
      "External id": 6600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020692.314, "dur": 4.488,
    "args": {
      "External id": 6601,"Record function id": 0, "Ev Idx": 6088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020693.676, "dur": 2.323,
    "args": {
      "External id": 6602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020694.353, "dur": 1.121,
    "args": {
      "External id": 6603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020694.733, "dur": 0.580,
    "args": {
      "External id": 6604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020700.217, "dur": 4.102,
    "args": {
      "External id": 6605,"Record function id": 0, "Ev Idx": 6092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020701.500, "dur": 2.141,
    "args": {
      "External id": 6606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020702.277, "dur": 0.850,
    "args": {
      "External id": 6607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020702.652, "dur": 0.345,
    "args": {
      "External id": 6608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020707.929, "dur": 4.614,
    "args": {
      "External id": 6609,"Record function id": 0, "Ev Idx": 6096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020709.508, "dur": 2.326,
    "args": {
      "External id": 6610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020710.117, "dur": 1.180,
    "args": {
      "External id": 6611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020710.489, "dur": 0.676,
    "args": {
      "External id": 6612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020716.334, "dur": 4.909,
    "args": {
      "External id": 6613,"Record function id": 0, "Ev Idx": 6100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020718.067, "dur": 2.361,
    "args": {
      "External id": 6614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 6101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020718.799, "dur": 1.127,
    "args": {
      "External id": 6615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 6102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020719.201, "dur": 0.600,
    "args": {
      "External id": 6616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 6103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020724.950, "dur": 5.821,
    "args": {
      "External id": 6617,"Record function id": 0, "Ev Idx": 6104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020727.014, "dur": 3.046,
    "args": {
      "External id": 6618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020727.944, "dur": 1.490,
    "args": {
      "External id": 6619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020728.340, "dur": 0.965,
    "args": {
      "External id": 6620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020734.148, "dur": 5.846,
    "args": {
      "External id": 6621,"Record function id": 0, "Ev Idx": 6108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020736.042, "dur": 3.101,
    "args": {
      "External id": 6622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020736.954, "dur": 1.620,
    "args": {
      "External id": 6623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020737.318, "dur": 1.124,
    "args": {
      "External id": 6624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020743.764, "dur": 5.036,
    "args": {
      "External id": 6625,"Record function id": 0, "Ev Idx": 6112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020745.756, "dur": 2.210,
    "args": {
      "External id": 6626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020746.539, "dur": 0.972,
    "args": {
      "External id": 6627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020746.920, "dur": 0.449,
    "args": {
      "External id": 6628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020752.421, "dur": 5.154,
    "args": {
      "External id": 6629,"Record function id": 0, "Ev Idx": 6116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020754.070, "dur": 2.822,
    "args": {
      "External id": 6630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020754.875, "dur": 1.503,
    "args": {
      "External id": 6631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020755.379, "dur": 0.854,
    "args": {
      "External id": 6632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020761.407, "dur": 4.556,
    "args": {
      "External id": 6633,"Record function id": 0, "Ev Idx": 6120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020762.764, "dur": 2.511,
    "args": {
      "External id": 6634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020763.632, "dur": 1.085,
    "args": {
      "External id": 6635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020764.011, "dur": 0.573,
    "args": {
      "External id": 6636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[12]], "Ev Idx": 6123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020769.170, "dur": 4.939,
    "args": {
      "External id": 6637,"Record function id": 0, "Ev Idx": 6124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020770.951, "dur": 2.472,
    "args": {
      "External id": 6638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020771.721, "dur": 1.245,
    "args": {
      "External id": 6639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020772.084, "dur": 0.740,
    "args": {
      "External id": 6640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020777.894, "dur": 5.610,
    "args": {
      "External id": 6641,"Record function id": 0, "Ev Idx": 6128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020779.586, "dur": 3.023,
    "args": {
      "External id": 6642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020780.263, "dur": 1.871,
    "args": {
      "External id": 6643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020780.653, "dur": 1.351,
    "args": {
      "External id": 6644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020786.922, "dur": 4.305,
    "args": {
      "External id": 6645,"Record function id": 0, "Ev Idx": 6132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020788.290, "dur": 2.104,
    "args": {
      "External id": 6646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020788.847, "dur": 1.063,
    "args": {
      "External id": 6647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020789.240, "dur": 0.542,
    "args": {
      "External id": 6648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020794.716, "dur": 4.986,
    "args": {
      "External id": 6649,"Record function id": 0, "Ev Idx": 6136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020796.109, "dur": 2.767,
    "args": {
      "External id": 6650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020796.842, "dur": 1.408,
    "args": {
      "External id": 6651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020797.406, "dur": 0.716,
    "args": {
      "External id": 6652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020803.128, "dur": 4.602,
    "args": {
      "External id": 6653,"Record function id": 0, "Ev Idx": 6140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508020804.563, "dur": 2.325,
    "args": {
      "External id": 6654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020805.354, "dur": 1.210,
    "args": {
      "External id": 6655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508020805.880, "dur": 0.555,
    "args": {
      "External id": 6656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021031.364, "dur": 15.852,
    "args": {
      "External id": 6657,"Record function id": 0, "Ev Idx": 6144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021039.570, "dur": 6.424,
    "args": {
      "External id": 6658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 6145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021041.424, "dur": 3.556,
    "args": {
      "External id": 6659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 6146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021042.504, "dur": 2.144,
    "args": {
      "External id": 6660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[4, 4, 1]], "Input Dims": [[384, 1, 4]], "Ev Idx": 6147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021054.419, "dur": 4.918,
    "args": {
      "External id": 6661,"Record function id": 0, "Ev Idx": 6148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021056.104, "dur": 2.381,
    "args": {
      "External id": 6662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021056.716, "dur": 1.154,
    "args": {
      "External id": 6663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021057.197, "dur": 0.548,
    "args": {
      "External id": 6664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021063.673, "dur": 5.060,
    "args": {
      "External id": 6665,"Record function id": 0, "Ev Idx": 6152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021065.569, "dur": 2.326,
    "args": {
      "External id": 6666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 6153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021066.284, "dur": 1.111,
    "args": {
      "External id": 6667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 6154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021066.753, "dur": 0.501,
    "args": {
      "External id": 6668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[256, 1]], "Input Dims": [[384, 256]], "Ev Idx": 6155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021072.284, "dur": 14.751,
    "args": {
      "External id": 6669,"Record function id": 0, "Ev Idx": 6156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021074.470, "dur": 5.063,
    "args": {
      "External id": 6670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021077.479, "dur": 1.532,
    "args": {
      "External id": 6671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021077.823, "dur": 1.045,
    "args": {
      "External id": 6672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021090.834, "dur": 5.038,
    "args": {
      "External id": 6673,"Record function id": 0, "Ev Idx": 6160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021092.597, "dur": 2.429,
    "args": {
      "External id": 6674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021093.279, "dur": 1.217,
    "args": {
      "External id": 6675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021093.618, "dur": 0.758,
    "args": {
      "External id": 6676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[16, 1]], "Input Dims": [[12, 16]], "Ev Idx": 6163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021099.603, "dur": 4.885,
    "args": {
      "External id": 6677,"Record function id": 0, "Ev Idx": 6164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021101.296, "dur": 2.506,
    "args": {
      "External id": 6678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021101.878, "dur": 1.410,
    "args": {
      "External id": 6679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021102.337, "dur": 0.808,
    "args": {
      "External id": 6680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[768, 1]], "Input Dims": [[384, 768]], "Ev Idx": 6167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021108.036, "dur": 5.051,
    "args": {
      "External id": 6681,"Record function id": 0, "Ev Idx": 6168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021109.908, "dur": 2.466,
    "args": {
      "External id": 6682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 6169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021110.456, "dur": 1.406,
    "args": {
      "External id": 6683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 6170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021111.065, "dur": 0.670,
    "args": {
      "External id": 6684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[32, 1]], "Input Dims": [[33048, 32]], "Ev Idx": 6171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021117.124, "dur": 5.872,
    "args": {
      "External id": 6685,"Record function id": 0, "Ev Idx": 6172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021118.937, "dur": 3.241,
    "args": {
      "External id": 6686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021119.812, "dur": 1.416,
    "args": {
      "External id": 6687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021120.238, "dur": 0.848,
    "args": {
      "External id": 6688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021126.648, "dur": 5.647,
    "args": {
      "External id": 6689,"Record function id": 0, "Ev Idx": 6176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021128.486, "dur": 2.992,
    "args": {
      "External id": 6690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021129.487, "dur": 1.322,
    "args": {
      "External id": 6691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021129.913, "dur": 0.770,
    "args": {
      "External id": 6692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[768, 384]], "Ev Idx": 6179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021135.848, "dur": 4.693,
    "args": {
      "External id": 6693,"Record function id": 0, "Ev Idx": 6180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021137.419, "dur": 2.449,
    "args": {
      "External id": 6694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021138.068, "dur": 1.285,
    "args": {
      "External id": 6695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021138.521, "dur": 0.708,
    "args": {
      "External id": 6696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021144.522, "dur": 4.813,
    "args": {
      "External id": 6697,"Record function id": 0, "Ev Idx": 6184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021145.968, "dur": 2.676,
    "args": {
      "External id": 6698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021146.920, "dur": 1.214,
    "args": {
      "External id": 6699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021147.307, "dur": 0.689,
    "args": {
      "External id": 6700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[384]], "Ev Idx": 6187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021152.773, "dur": 4.925,
    "args": {
      "External id": 6701,"Record function id": 0, "Ev Idx": 6188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021154.217, "dur": 2.648,
    "args": {
      "External id": 6702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021154.879, "dur": 1.267,
    "args": {
      "External id": 6703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021155.258, "dur": 0.768,
    "args": {
      "External id": 6704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021161.447, "dur": 4.303,
    "args": {
      "External id": 6705,"Record function id": 0, "Ev Idx": 6192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021162.824, "dur": 2.241,
    "args": {
      "External id": 6706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021163.540, "dur": 1.053,
    "args": {
      "External id": 6707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021163.877, "dur": 0.589,
    "args": {
      "External id": 6708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[384, 384]], "Ev Idx": 6195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021169.135, "dur": 4.841,
    "args": {
      "External id": 6709,"Record function id": 0, "Ev Idx": 6196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021170.632, "dur": 2.535,
    "args": {
      "External id": 6710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021171.334, "dur": 1.223,
    "args": {
      "External id": 6711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021171.804, "dur": 0.585,
    "args": {
      "External id": 6712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[32]], "Ev Idx": 6199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508021177.799, "dur": 181.059,
    "args": {
      "External id": 6713,"Record function id": 0, "Sequence number": 4698, "Fwd thread id": 1, "Ev Idx": 6200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 14565, "tid": 14640,
    "ts": 1290508021179.932, "dur": 147.792,
    "args": {
      "External id": 6714,"Sequence number": 4698, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["", "float", "", "", "", ""], "Input Strides": [[], [393216, 384, 1], [], [], [], []], "Input Dims": [[], [8, 1024, 384], [], [], [], []], "Ev Idx": 6201
    }
  },
  {
    "ph": "f", "id": 16, "pid": 14565, "tid": 14640, "ts": 1290508021179.932,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 14565, "tid": 14640,
    "ts": 1290508021251.078, "dur": 65.626,
    "args": {
      "External id": 6715,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 14565, "tid": 14640,
    "ts": 1290508021334.354, "dur": 21.761,
    "args": {
      "External id": 6716,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[384, 1], [384, 1], []], "Input Dims": [[12288, 384], [12288, 384], []], "Ev Idx": 6203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021363.989, "dur": 5.472,
    "args": {
      "External id": 6717,"Record function id": 0, "Ev Idx": 6204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 14565, "tid": 14640,
    "ts": 1290508021365.597, "dur": 2.988,
    "args": {
      "External id": 6718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 6205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021366.560, "dur": 1.710,
    "args": {
      "External id": 6719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 6206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14640,
    "ts": 1290508021367.201, "dur": 0.913,
    "args": {
      "External id": 6720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[384, 1]], "Input Dims": [[12288, 384]], "Ev Idx": 6207
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290507962190.045,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6208
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290507970473.888,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6209
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290507980634.412,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6210
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290507989857.725,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6211
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290507997157.391,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6212
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290508006965.336,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6213
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290508013016.642,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6214
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14640,
    "ts": 1290508021340.060,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222848, "Bytes": 18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6215
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#3", "pid": 14565, "tid": 14565,
    "ts": 1290507954196.049, "dur": 8138.299,
    "args": {
      "External id": 1,"Record function id": 0, "Ev Idx": 6216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507954788.152, "dur": 29.341,
    "args": {
      "External id": 2,"Record function id": 0, "Ev Idx": 6217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290507954820.130, "dur": 2171.088,
    "args": {
      "External id": 3,"Record function id": 0, "Ev Idx": 6218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507954831.405, "dur": 3.983,
    "args": {
      "External id": 4,"Record function id": 0, "Ev Idx": 6219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507954848.694, "dur": 10.680,
    "args": {
      "External id": 5,"Record function id": 0, "Ev Idx": 6220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507954896.002, "dur": 198.553,
    "args": {
      "External id": 6,"Sequence number": 4684, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6221
    }
  },
  {
    "ph": "s", "id": 2, "pid": 14565, "tid": 14565, "ts": 1290507954896.002,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290507954912.376, "dur": 137.292,
    "args": {
      "External id": 7,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507954970.860, "dur": 26.661,
    "args": {
      "External id": 8,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507955120.649, "dur": 79.663,
    "args": {
      "External id": 9,"Record function id": 0, "Ev Idx": 6224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290507955201.796, "dur": 1785.780,
    "args": {
      "External id": 10,"Record function id": 0, "Ev Idx": 6225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507955206.698, "dur": 20.765,
    "args": {
      "External id": 11,"Record function id": 0, "Ev Idx": 6226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507955241.330, "dur": 9.068,
    "args": {
      "External id": 12,"Record function id": 0, "Ev Idx": 6227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507955374.762, "dur": 1547.398,
    "args": {
      "External id": 13,"Sequence number": 4685, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6228
    }
  },
  {
    "ph": "s", "id": 1, "pid": 14565, "tid": 14565, "ts": 1290507955374.762,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290507955391.350, "dur": 1122.798,
    "args": {
      "External id": 14,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507955487.699, "dur": 67.283,
    "args": {
      "External id": 15,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507955494.102, "dur": 27.838,
    "args": {
      "External id": 16,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507955523.085, "dur": 7.803,
    "args": {
      "External id": 17,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507955531.758, "dur": 6.363,
    "args": {
      "External id": 18,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507955538.915, "dur": 6.563,
    "args": {
      "External id": 19,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507955546.437, "dur": 7.106,
    "args": {
      "External id": 20,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507956956.041, "dur": 3.480,
    "args": {
      "External id": 21,"Sequence number": 4686, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507956957.343, "dur": 1.983,
    "args": {
      "External id": 22,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507956964.313, "dur": 2.426,
    "args": {
      "External id": 23,"Sequence number": 4686, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507956965.220, "dur": 1.348,
    "args": {
      "External id": 24,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507956972.490, "dur": 1.205,
    "args": {
      "External id": 25,"Sequence number": 4686, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507956972.972, "dur": 0.550,
    "args": {
      "External id": 26,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290507957040.545, "dur": 51.974,
    "args": {
      "External id": 27,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290507957042.955, "dur": 18.859,
    "args": {
      "External id": 28,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290507957051.873, "dur": 9.339,
    "args": {
      "External id": 29,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290507957063.701, "dur": 28.417,
    "args": {
      "External id": 30,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6245
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#4", "pid": 14565, "tid": 14565,
    "ts": 1290507962355.604, "dur": 8272.926,
    "args": {
      "External id": 31,"Record function id": 0, "Ev Idx": 6246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507962956.028, "dur": 23.561,
    "args": {
      "External id": 32,"Record function id": 0, "Ev Idx": 6247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290507962981.353, "dur": 1721.405,
    "args": {
      "External id": 33,"Record function id": 0, "Ev Idx": 6248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507962989.497, "dur": 3.316,
    "args": {
      "External id": 34,"Record function id": 0, "Ev Idx": 6249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507963002.663, "dur": 10.240,
    "args": {
      "External id": 35,"Record function id": 0, "Ev Idx": 6250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507963029.515, "dur": 180.571,
    "args": {
      "External id": 36,"Sequence number": 4686, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6251
    }
  },
  {
    "ph": "s", "id": 4, "pid": 14565, "tid": 14565, "ts": 1290507963029.515,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290507963042.630, "dur": 128.758,
    "args": {
      "External id": 37,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507963095.559, "dur": 25.243,
    "args": {
      "External id": 38,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507963234.053, "dur": 61.162,
    "args": {
      "External id": 39,"Record function id": 0, "Ev Idx": 6254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290507963296.297, "dur": 1402.922,
    "args": {
      "External id": 40,"Record function id": 0, "Ev Idx": 6255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507963300.775, "dur": 19.758,
    "args": {
      "External id": 41,"Record function id": 0, "Ev Idx": 6256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507963330.484, "dur": 7.941,
    "args": {
      "External id": 42,"Record function id": 0, "Ev Idx": 6257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507963446.101, "dur": 1204.048,
    "args": {
      "External id": 43,"Sequence number": 4687, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6258
    }
  },
  {
    "ph": "s", "id": 3, "pid": 14565, "tid": 14565, "ts": 1290507963446.101,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290507963462.525, "dur": 886.543,
    "args": {
      "External id": 44,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507963555.364, "dur": 64.891,
    "args": {
      "External id": 45,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507963560.754, "dur": 28.183,
    "args": {
      "External id": 46,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507963590.247, "dur": 7.532,
    "args": {
      "External id": 47,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507963598.637, "dur": 6.018,
    "args": {
      "External id": 48,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507963605.542, "dur": 6.439,
    "args": {
      "External id": 49,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507963613.010, "dur": 5.793,
    "args": {
      "External id": 50,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507964671.490, "dur": 3.759,
    "args": {
      "External id": 51,"Sequence number": 4688, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507964672.567, "dur": 2.388,
    "args": {
      "External id": 52,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507964679.047, "dur": 1.502,
    "args": {
      "External id": 53,"Sequence number": 4688, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507964679.813, "dur": 0.574,
    "args": {
      "External id": 54,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507964685.091, "dur": 1.453,
    "args": {
      "External id": 55,"Sequence number": 4688, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507964685.773, "dur": 0.625,
    "args": {
      "External id": 56,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290507964736.965, "dur": 34.515,
    "args": {
      "External id": 57,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290507964738.112, "dur": 12.103,
    "args": {
      "External id": 58,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290507964740.882, "dur": 8.815,
    "args": {
      "External id": 59,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290507964751.597, "dur": 19.513,
    "args": {
      "External id": 60,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6275
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#5", "pid": 14565, "tid": 14565,
    "ts": 1290507970659.156, "dur": 10088.462,
    "args": {
      "External id": 61,"Record function id": 0, "Ev Idx": 6276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507971454.741, "dur": 34.258,
    "args": {
      "External id": 62,"Record function id": 0, "Ev Idx": 6277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290507971491.417, "dur": 2795.032,
    "args": {
      "External id": 63,"Record function id": 0, "Ev Idx": 6278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507971501.539, "dur": 4.055,
    "args": {
      "External id": 64,"Record function id": 0, "Ev Idx": 6279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507971518.724, "dur": 14.005,
    "args": {
      "External id": 65,"Record function id": 0, "Ev Idx": 6280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507971553.261, "dur": 200.226,
    "args": {
      "External id": 66,"Sequence number": 4688, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6281
    }
  },
  {
    "ph": "s", "id": 6, "pid": 14565, "tid": 14565, "ts": 1290507971553.261,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290507971568.945, "dur": 140.735,
    "args": {
      "External id": 67,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507971619.806, "dur": 28.983,
    "args": {
      "External id": 68,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507971782.233, "dur": 100.897,
    "args": {
      "External id": 69,"Record function id": 0, "Ev Idx": 6284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290507971884.572, "dur": 2396.205,
    "args": {
      "External id": 70,"Record function id": 0, "Ev Idx": 6285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507971889.872, "dur": 30.046,
    "args": {
      "External id": 71,"Record function id": 0, "Ev Idx": 6286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507971934.776, "dur": 11.367,
    "args": {
      "External id": 72,"Record function id": 0, "Ev Idx": 6287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507972110.763, "dur": 2082.838,
    "args": {
      "External id": 73,"Sequence number": 4689, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6288
    }
  },
  {
    "ph": "s", "id": 5, "pid": 14565, "tid": 14565, "ts": 1290507972110.763,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290507972129.162, "dur": 1555.363,
    "args": {
      "External id": 74,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507972336.695, "dur": 84.712,
    "args": {
      "External id": 75,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507972344.182, "dur": 37.439,
    "args": {
      "External id": 76,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507972383.881, "dur": 9.044,
    "args": {
      "External id": 77,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507972394.348, "dur": 6.961,
    "args": {
      "External id": 78,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507972402.403, "dur": 7.920,
    "args": {
      "External id": 79,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507972412.310, "dur": 7.216,
    "args": {
      "External id": 80,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507974231.295, "dur": 9.613,
    "args": {
      "External id": 81,"Sequence number": 4690, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507974232.930, "dur": 7.601,
    "args": {
      "External id": 82,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507974246.365, "dur": 1.850,
    "args": {
      "External id": 83,"Sequence number": 4690, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507974247.447, "dur": 0.619,
    "args": {
      "External id": 84,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507974253.592, "dur": 1.665,
    "args": {
      "External id": 85,"Sequence number": 4690, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507974254.073, "dur": 1.030,
    "args": {
      "External id": 86,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290507974334.253, "dur": 47.685,
    "args": {
      "External id": 87,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290507974335.792, "dur": 15.958,
    "args": {
      "External id": 88,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290507974339.244, "dur": 12.043,
    "args": {
      "External id": 89,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290507974353.244, "dur": 28.200,
    "args": {
      "External id": 90,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6305
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#6", "pid": 14565, "tid": 14565,
    "ts": 1290507980771.972, "dur": 9190.867,
    "args": {
      "External id": 91,"Record function id": 0, "Ev Idx": 6306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507981439.038, "dur": 24.188,
    "args": {
      "External id": 92,"Record function id": 0, "Ev Idx": 6307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290507981464.864, "dur": 1981.785,
    "args": {
      "External id": 93,"Record function id": 0, "Ev Idx": 6308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507981473.508, "dur": 4.364,
    "args": {
      "External id": 94,"Record function id": 0, "Ev Idx": 6309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507981490.062, "dur": 14.066,
    "args": {
      "External id": 95,"Record function id": 0, "Ev Idx": 6310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507981523.431, "dur": 190.656,
    "args": {
      "External id": 96,"Sequence number": 4690, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6311
    }
  },
  {
    "ph": "s", "id": 8, "pid": 14565, "tid": 14565, "ts": 1290507981523.431,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290507981540.121, "dur": 133.203,
    "args": {
      "External id": 97,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507981587.596, "dur": 25.179,
    "args": {
      "External id": 98,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507981740.436, "dur": 74.551,
    "args": {
      "External id": 99,"Record function id": 0, "Ev Idx": 6314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290507981816.149, "dur": 1627.123,
    "args": {
      "External id": 100,"Record function id": 0, "Ev Idx": 6315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507981821.041, "dur": 21.573,
    "args": {
      "External id": 101,"Record function id": 0, "Ev Idx": 6316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507981855.385, "dur": 9.872,
    "args": {
      "External id": 102,"Record function id": 0, "Ev Idx": 6317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507982022.358, "dur": 1362.200,
    "args": {
      "External id": 103,"Sequence number": 4691, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6318
    }
  },
  {
    "ph": "s", "id": 7, "pid": 14565, "tid": 14565, "ts": 1290507982022.358,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290507982043.196, "dur": 962.114,
    "args": {
      "External id": 104,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507982151.473, "dur": 62.733,
    "args": {
      "External id": 105,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507982156.838, "dur": 28.306,
    "args": {
      "External id": 106,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507982186.300, "dur": 6.797,
    "args": {
      "External id": 107,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507982193.968, "dur": 5.834,
    "args": {
      "External id": 108,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507982200.586, "dur": 5.886,
    "args": {
      "External id": 109,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507982207.282, "dur": 5.526,
    "args": {
      "External id": 110,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507983411.648, "dur": 4.515,
    "args": {
      "External id": 111,"Sequence number": 4692, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507983412.838, "dur": 3.019,
    "args": {
      "External id": 112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507983420.582, "dur": 1.543,
    "args": {
      "External id": 113,"Sequence number": 4692, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507983421.431, "dur": 0.542,
    "args": {
      "External id": 114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507983427.129, "dur": 1.657,
    "args": {
      "External id": 115,"Sequence number": 4692, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507983427.882, "dur": 0.723,
    "args": {
      "External id": 116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290507983483.872, "dur": 46.203,
    "args": {
      "External id": 117,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290507983485.251, "dur": 15.361,
    "args": {
      "External id": 118,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290507983488.705, "dur": 11.323,
    "args": {
      "External id": 119,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290507983502.014, "dur": 27.630,
    "args": {
      "External id": 120,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6335
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#7", "pid": 14565, "tid": 14565,
    "ts": 1290507989982.898, "dur": 7288.016,
    "args": {
      "External id": 121,"Record function id": 0, "Ev Idx": 6336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507990665.744, "dur": 21.406,
    "args": {
      "External id": 122,"Record function id": 0, "Ev Idx": 6337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290507990688.705, "dur": 1774.179,
    "args": {
      "External id": 123,"Record function id": 0, "Ev Idx": 6338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507990696.532, "dur": 3.457,
    "args": {
      "External id": 124,"Record function id": 0, "Ev Idx": 6339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507990709.827, "dur": 10.426,
    "args": {
      "External id": 125,"Record function id": 0, "Ev Idx": 6340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507990736.630, "dur": 165.946,
    "args": {
      "External id": 126,"Sequence number": 4692, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6341
    }
  },
  {
    "ph": "s", "id": 10, "pid": 14565, "tid": 14565, "ts": 1290507990736.630,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290507990750.531, "dur": 114.692,
    "args": {
      "External id": 127,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507990790.696, "dur": 22.704,
    "args": {
      "External id": 128,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507990927.712, "dur": 55.570,
    "args": {
      "External id": 129,"Record function id": 0, "Ev Idx": 6344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290507990984.623, "dur": 1475.176,
    "args": {
      "External id": 130,"Record function id": 0, "Ev Idx": 6345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507990988.971, "dur": 19.929,
    "args": {
      "External id": 131,"Record function id": 0, "Ev Idx": 6346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507991019.664, "dur": 8.654,
    "args": {
      "External id": 132,"Record function id": 0, "Ev Idx": 6347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507991179.389, "dur": 1224.688,
    "args": {
      "External id": 133,"Sequence number": 4693, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6348
    }
  },
  {
    "ph": "s", "id": 9, "pid": 14565, "tid": 14565, "ts": 1290507991179.389,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290507991195.971, "dur": 931.569,
    "args": {
      "External id": 134,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507991286.837, "dur": 63.564,
    "args": {
      "External id": 135,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507991291.796, "dur": 27.643,
    "args": {
      "External id": 136,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507991320.611, "dur": 7.651,
    "args": {
      "External id": 137,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507991329.169, "dur": 6.158,
    "args": {
      "External id": 138,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507991336.298, "dur": 6.251,
    "args": {
      "External id": 139,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507991343.369, "dur": 5.648,
    "args": {
      "External id": 140,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507992425.565, "dur": 8.941,
    "args": {
      "External id": 141,"Sequence number": 4694, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507992426.829, "dur": 7.465,
    "args": {
      "External id": 142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507992438.542, "dur": 1.605,
    "args": {
      "External id": 143,"Sequence number": 4694, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507992439.461, "dur": 0.530,
    "args": {
      "External id": 144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507992445.176, "dur": 1.580,
    "args": {
      "External id": 145,"Sequence number": 4694, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507992445.683, "dur": 0.943,
    "args": {
      "External id": 146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290507992497.459, "dur": 35.524,
    "args": {
      "External id": 147,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290507992498.658, "dur": 12.397,
    "args": {
      "External id": 148,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290507992501.822, "dur": 8.656,
    "args": {
      "External id": 149,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290507992512.470, "dur": 20.085,
    "args": {
      "External id": 150,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6365
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#8", "pid": 14565, "tid": 14565,
    "ts": 1290507997292.419, "dur": 9779.014,
    "args": {
      "External id": 151,"Record function id": 0, "Ev Idx": 6366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507997856.629, "dur": 20.667,
    "args": {
      "External id": 152,"Record function id": 0, "Ev Idx": 6367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290507997879.141, "dur": 1658.364,
    "args": {
      "External id": 153,"Record function id": 0, "Ev Idx": 6368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507997887.107, "dur": 3.350,
    "args": {
      "External id": 154,"Record function id": 0, "Ev Idx": 6369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507997901.238, "dur": 10.321,
    "args": {
      "External id": 155,"Record function id": 0, "Ev Idx": 6370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507997928.328, "dur": 177.658,
    "args": {
      "External id": 156,"Sequence number": 4694, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6371
    }
  },
  {
    "ph": "s", "id": 12, "pid": 14565, "tid": 14565, "ts": 1290507997928.328,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290507997942.450, "dur": 117.300,
    "args": {
      "External id": 157,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507997984.461, "dur": 25.267,
    "args": {
      "External id": 158,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290507998131.389, "dur": 53.596,
    "args": {
      "External id": 159,"Record function id": 0, "Ev Idx": 6374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290507998186.185, "dur": 1348.210,
    "args": {
      "External id": 160,"Record function id": 0, "Ev Idx": 6375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290507998191.074, "dur": 19.568,
    "args": {
      "External id": 161,"Record function id": 0, "Ev Idx": 6376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290507998221.205, "dur": 8.678,
    "args": {
      "External id": 162,"Record function id": 0, "Ev Idx": 6377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290507998340.217, "dur": 1143.031,
    "args": {
      "External id": 163,"Sequence number": 4695, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6378
    }
  },
  {
    "ph": "s", "id": 11, "pid": 14565, "tid": 14565, "ts": 1290507998340.217,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290507998355.774, "dur": 837.120,
    "args": {
      "External id": 164,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507998446.369, "dur": 66.279,
    "args": {
      "External id": 165,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507998451.437, "dur": 28.832,
    "args": {
      "External id": 166,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507998481.444, "dur": 7.567,
    "args": {
      "External id": 167,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507998490.109, "dur": 6.483,
    "args": {
      "External id": 168,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507998497.505, "dur": 7.157,
    "args": {
      "External id": 169,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290507998505.520, "dur": 5.545,
    "args": {
      "External id": 170,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507999505.488, "dur": 3.574,
    "args": {
      "External id": 171,"Sequence number": 4696, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507999506.576, "dur": 2.300,
    "args": {
      "External id": 172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507999513.119, "dur": 1.630,
    "args": {
      "External id": 173,"Sequence number": 4696, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507999514.028, "dur": 0.568,
    "args": {
      "External id": 174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290507999519.557, "dur": 1.340,
    "args": {
      "External id": 175,"Sequence number": 4696, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290507999520.249, "dur": 0.524,
    "args": {
      "External id": 176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290507999574.540, "dur": 36.047,
    "args": {
      "External id": 177,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290507999575.810, "dur": 12.824,
    "args": {
      "External id": 178,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290507999579.129, "dur": 8.948,
    "args": {
      "External id": 179,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290507999589.994, "dur": 20.197,
    "args": {
      "External id": 180,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6395
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#9", "pid": 14565, "tid": 14565,
    "ts": 1290508007098.280, "dur": 6056.186,
    "args": {
      "External id": 181,"Record function id": 0, "Ev Idx": 6396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290508007681.672, "dur": 19.285,
    "args": {
      "External id": 182,"Record function id": 0, "Ev Idx": 6397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290508007702.456, "dur": 1643.800,
    "args": {
      "External id": 183,"Record function id": 0, "Ev Idx": 6398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290508007710.383, "dur": 3.258,
    "args": {
      "External id": 184,"Record function id": 0, "Ev Idx": 6399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290508007723.429, "dur": 10.475,
    "args": {
      "External id": 185,"Record function id": 0, "Ev Idx": 6400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290508007750.180, "dur": 158.552,
    "args": {
      "External id": 186,"Sequence number": 4696, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6401
    }
  },
  {
    "ph": "s", "id": 14, "pid": 14565, "tid": 14565, "ts": 1290508007750.180,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290508007762.655, "dur": 108.763,
    "args": {
      "External id": 187,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508007801.325, "dur": 21.765,
    "args": {
      "External id": 188,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290508007932.711, "dur": 50.967,
    "args": {
      "External id": 189,"Record function id": 0, "Ev Idx": 6404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290508007984.942, "dur": 1358.036,
    "args": {
      "External id": 190,"Record function id": 0, "Ev Idx": 6405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290508007989.263, "dur": 17.600,
    "args": {
      "External id": 191,"Record function id": 0, "Ev Idx": 6406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290508008017.908, "dur": 8.100,
    "args": {
      "External id": 192,"Record function id": 0, "Ev Idx": 6407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290508008138.178, "dur": 1145.061,
    "args": {
      "External id": 193,"Sequence number": 4697, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6408
    }
  },
  {
    "ph": "s", "id": 13, "pid": 14565, "tid": 14565, "ts": 1290508008138.178,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290508008153.569, "dur": 785.028,
    "args": {
      "External id": 194,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508008238.291, "dur": 64.207,
    "args": {
      "External id": 195,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508008243.484, "dur": 27.128,
    "args": {
      "External id": 196,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508008271.971, "dur": 7.337,
    "args": {
      "External id": 197,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508008280.548, "dur": 6.077,
    "args": {
      "External id": 198,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508008287.708, "dur": 6.583,
    "args": {
      "External id": 199,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508008295.316, "dur": 5.720,
    "args": {
      "External id": 200,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290508009308.137, "dur": 7.409,
    "args": {
      "External id": 201,"Sequence number": 4698, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290508009309.011, "dur": 6.312,
    "args": {
      "External id": 202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290508009320.042, "dur": 1.449,
    "args": {
      "External id": 203,"Sequence number": 4698, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290508009320.947, "dur": 0.393,
    "args": {
      "External id": 204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290508009326.317, "dur": 1.297,
    "args": {
      "External id": 205,"Sequence number": 4698, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290508009327.002, "dur": 0.473,
    "args": {
      "External id": 206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290508009382.933, "dur": 42.273,
    "args": {
      "External id": 207,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290508009384.251, "dur": 12.796,
    "args": {
      "External id": 208,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290508009387.319, "dur": 9.216,
    "args": {
      "External id": 209,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290508009398.474, "dur": 26.291,
    "args": {
      "External id": 210,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6425
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#10", "pid": 14565, "tid": 14565,
    "ts": 1290508013177.776, "dur": 8275.684,
    "args": {
      "External id": 211,"Record function id": 0, "Ev Idx": 6426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290508013752.591, "dur": 20.281,
    "args": {
      "External id": 212,"Record function id": 0, "Ev Idx": 6427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 14565, "tid": 14565,
    "ts": 1290508013774.692, "dur": 1853.129,
    "args": {
      "External id": 213,"Record function id": 0, "Ev Idx": 6428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290508013782.420, "dur": 3.070,
    "args": {
      "External id": 214,"Record function id": 0, "Ev Idx": 6429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290508013797.782, "dur": 9.999,
    "args": {
      "External id": 215,"Record function id": 0, "Ev Idx": 6430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290508013841.018, "dur": 184.832,
    "args": {
      "External id": 216,"Sequence number": 4698, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", ""], "Input type": ["long int", "long int", "float", "long int", "long int", "long int", "long int"], "Input Strides": [[1024, 1], [1024, 1], [384, 1], [1], [3, 1], [8, 1], [8, 1]], "Input Dims": [[8, 1024], [8, 1024], [12288, 384], [12288], [2, 3], [2, 8], [2, 8]], "Ev Idx": 6431
    }
  },
  {
    "ph": "s", "id": 16, "pid": 14565, "tid": 14565, "ts": 1290508013841.018,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 14565, "tid": 14565,
    "ts": 1290508013855.352, "dur": 130.861,
    "args": {
      "External id": 217,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508013904.138, "dur": 29.679,
    "args": {
      "External id": 218,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1024, 1]], [[1024, 1]], []], "Input Dims": [[[8, 1024]], [[8, 1024]], []], "Ev Idx": 6433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 14565, "tid": 14565,
    "ts": 1290508014050.613, "dur": 60.036,
    "args": {
      "External id": 219,"Record function id": 0, "Ev Idx": 6434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/0", "pid": 14565, "tid": 14565,
    "ts": 1290508014111.932, "dur": 1511.740,
    "args": {
      "External id": 220,"Record function id": 0, "Ev Idx": 6435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Pregraph bytecode", "pid": 14565, "tid": 14565,
    "ts": 1290508014116.476, "dur": 18.947,
    "args": {
      "External id": 221,"Record function id": 0, "Ev Idx": 6436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AOTDispatcher Runtime Wrapper Prologue", "pid": 14565, "tid": 14565,
    "ts": 1290508014145.043, "dur": 8.083,
    "args": {
      "External id": 222,"Record function id": 0, "Ev Idx": 6437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 14565, "tid": 14565,
    "ts": 1290508014259.074, "dur": 1311.145,
    "args": {
      "External id": 223,"Sequence number": 4699, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "int", "int", "int", "int", "int", "int", "int", "int", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "c10::BFloat16", "c10::BFloat16", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "float", "long int", "float", "float", "float"], "Input Strides": [[1], [393216, 384, 1], [384, 1], [32, 1], [32, 1], [1], [1], [64, 64, 8, 1], [8, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [8192, 8, 1], [32, 1], [256, 1], [1], [1], [256, 1], [1], [4, 4, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [384, 1], [32, 1], [32, 1], [1], [1], [16, 1], [1], [1], [384, 1], [1], [384, 1], [384, 1], [768, 1], [1], [1024, 1], [384, 1], [], []], "Input Dims": [[384], [8, 1024, 384], [768, 384], [1152, 32], [1152, 32], [32], [32], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33048, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [8, 1024, 8], [33644, 32], [384, 256], [384], [384], [384, 256], [384], [384, 1, 4], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [768, 384], [1152, 32], [1152, 32], [32], [32], [12, 16], [12], [1152], [384, 384], [384], [768, 384], [768, 384], [384, 768], [384], [8, 1024], [12288, 384], [], []], "Ev Idx": 6438
    }
  },
  {
    "ph": "s", "id": 15, "pid": 14565, "tid": 14565, "ts": 1290508014259.074,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 14565, "tid": 14565,
    "ts": 1290508014274.846, "dur": 961.878,
    "args": {
      "External id": 224,"Record function id": 0, "scope": "user_scope", "Ev Idx": 6439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508014433.212, "dur": 66.475,
    "args": {
      "External id": 225,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], [[8, 8, 1], [64, 64, 8, 1], [8, 8, 1], [64, 64, 8, 1], [1024, 1]], []], "Input Dims": [[[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], [[1, 1, 8], [1, 1, 8, 8], [1, 1, 8], [1, 1, 8, 8], [8, 1024]], []], "Ev Idx": 6440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508014438.641, "dur": 29.717,
    "args": {
      "External id": 226,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508014469.826, "dur": 7.509,
    "args": {
      "External id": 227,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508014478.325, "dur": 6.402,
    "args": {
      "External id": 228,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[8, 8, 1], [8, 8, 1], []], "Input Dims": [[1, 1, 8], [1, 1, 8], []], "Ev Idx": 6443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508014485.614, "dur": 5.837,
    "args": {
      "External id": 229,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Strides": [[64, 64, 8, 1], [64, 64, 8, 1], []], "Input Dims": [[1, 1, 8, 8], [1, 1, 8, 8], []], "Ev Idx": 6444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 14565, "tid": 14565,
    "ts": 1290508014492.551, "dur": 5.415,
    "args": {
      "External id": 230,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1024, 1], [1024, 1], []], "Input Dims": [[8, 1024], [8, 1024], []], "Ev Idx": 6445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290508015593.524, "dur": 2.664,
    "args": {
      "External id": 231,"Sequence number": 4700, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290508015594.328, "dur": 1.592,
    "args": {
      "External id": 232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290508015599.211, "dur": 1.516,
    "args": {
      "External id": 233,"Sequence number": 4700, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290508015600.057, "dur": 0.526,
    "args": {
      "External id": 234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 14565, "tid": 14565,
    "ts": 1290508015604.626, "dur": 1.758,
    "args": {
      "External id": 235,"Sequence number": 4700, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 14565, "tid": 14565,
    "ts": 1290508015605.260, "dur": 0.937,
    "args": {
      "External id": 236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 6451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 14565, "tid": 14565,
    "ts": 1290508015657.904, "dur": 32.500,
    "args": {
      "External id": 237,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 14565, "tid": 14565,
    "ts": 1290508015659.110, "dur": 9.939,
    "args": {
      "External id": 238,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 14565, "tid": 14565,
    "ts": 1290508015661.788, "dur": 6.813,
    "args": {
      "External id": 239,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 14565, "tid": 14565,
    "ts": 1290508015670.331, "dur": 19.694,
    "args": {
      "External id": 240,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 6455
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507954230.571,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6456
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507957058.909,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6457
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507962319.924,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6458
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507962383.108,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6459
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507964747.770,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6460
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507970602.229,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6461
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507970702.933,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6462
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507974348.376,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6463
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507980729.905,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6464
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507980804.849,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6465
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507983496.918,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6466
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507989947.837,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6467
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507990012.806,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6468
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507992508.552,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6469
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507997255.242,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6470
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507997321.784,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6471
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290507999586.039,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6472
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290508007057.040,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6473
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290508007126.595,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6474
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290508009394.306,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6475
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290508013137.242,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6476
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290508013209.917,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203347968, "Bytes": -18874368, "Device Id": 0, "Device Type": 1, "Addr": 139907385262080, "finished": false, "Ev Idx": 6477
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290508015666.623,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 203348480, "Bytes": 512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6478
    }
  },
  {
    "ph": "i", "cat": "cpu_instant_event", "s": "t", "name": "[memory]",
    "pid": 14565, "tid": 14565,
    "ts": 1290508021438.701,
    "args": {
      "Total Reserved": 2996830208, "Total Allocated": 222222336, "Bytes": -512, "Device Id": 0, "Device Type": 1, "Addr": 139913448120832, "finished": false, "Ev Idx": 6479
    }
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290507960958.996, "dur": 3.392,
    "args": {
      "External id": 8,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502528,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502528, "pid": 0, "tid": 7, "ts": 1290507960958.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507954979.120, "dur": 15.568,
    "args": {
      "External id": 8,      "cbid": 211, "correlation": 502528
    }
  },
  {
    "ph": "s", "id": 502528, "pid": 14565, "tid": 14565, "ts": 1290507954979.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507955002.314, "dur": 0.574,
    "args": {
      "External id": 6,      "cbid": 317, "correlation": 502536
    }
  },
  {
    "ph": "f", "id": 502536, "pid": 14565, "tid": 14565, "ts": 1290507955002.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290507960966.676, "dur": 0.832,
    "args": {
      "External id": 6,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502538,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502538, "pid": 0, "tid": 7, "ts": 1290507960966.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290507960967.668, "dur": 0.800,
    "args": {
      "External id": 6,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502538,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502538, "pid": 0, "tid": 7, "ts": 1290507960967.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290507960968.596, "dur": 2.112,
    "args": {
      "External id": 6,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502538,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502538, "pid": 0, "tid": 7, "ts": 1290507960968.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290507960970.836, "dur": 6.688,
    "args": {
      "External id": 6,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502538,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502538, "pid": 0, "tid": 7, "ts": 1290507960970.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290507960977.620, "dur": 11.424,
    "args": {
      "External id": 6,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502538,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502538, "pid": 0, "tid": 7, "ts": 1290507960977.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507955003.574, "dur": 15.342,
    "args": {
      "External id": 6,      "cbid": 311, "correlation": 502538
    }
  },
  {
    "ph": "s", "id": 502538, "pid": 14565, "tid": 14565, "ts": 1290507955003.574,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507960991.508, "dur": 1.344,
    "args": {
      "External id": 16,      "device": 0, "context": 1,      "stream": 7, "correlation": 502543,      "bytes": 32, "memory bandwidth (GB/s)": 0.023809523809523808
    }
  },
  {
    "ph": "f", "id": 502543, "pid": 0, "tid": 7, "ts": 1290507960991.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507955501.441, "dur": 17.764,
    "args": {
      "External id": 16,      "cbid": 41, "correlation": 502543
    }
  },
  {
    "ph": "s", "id": 502543, "pid": 14565, "tid": 14565, "ts": 1290507955501.441,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507960993.268, "dur": 1.280,
    "args": {
      "External id": 17,      "device": 0, "context": 1,      "stream": 7, "correlation": 502548,      "bytes": 256, "memory bandwidth (GB/s)": 0.2
    }
  },
  {
    "ph": "f", "id": 502548, "pid": 0, "tid": 7, "ts": 1290507960993.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507955524.587, "dur": 5.049,
    "args": {
      "External id": 17,      "cbid": 41, "correlation": 502548
    }
  },
  {
    "ph": "s", "id": 502548, "pid": 14565, "tid": 14565, "ts": 1290507955524.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507960994.964, "dur": 1.376,
    "args": {
      "External id": 18,      "device": 0, "context": 1,      "stream": 7, "correlation": 502553,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 502553, "pid": 0, "tid": 7, "ts": 1290507960994.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507955532.744, "dur": 4.377,
    "args": {
      "External id": 18,      "cbid": 41, "correlation": 502553
    }
  },
  {
    "ph": "s", "id": 502553, "pid": 14565, "tid": 14565, "ts": 1290507955532.744,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507960996.756, "dur": 1.312,
    "args": {
      "External id": 19,      "device": 0, "context": 1,      "stream": 7, "correlation": 502558,      "bytes": 256, "memory bandwidth (GB/s)": 0.1951219512195122
    }
  },
  {
    "ph": "f", "id": 502558, "pid": 0, "tid": 7, "ts": 1290507960996.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507955539.825, "dur": 4.552,
    "args": {
      "External id": 19,      "cbid": 41, "correlation": 502558
    }
  },
  {
    "ph": "s", "id": 502558, "pid": 14565, "tid": 14565, "ts": 1290507955539.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507960998.516, "dur": 2.048,
    "args": {
      "External id": 20,      "device": 0, "context": 1,      "stream": 7, "correlation": 502563,      "bytes": 65536, "memory bandwidth (GB/s)": 32
    }
  },
  {
    "ph": "f", "id": 502563, "pid": 0, "tid": 7, "ts": 1290507960998.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507955547.378, "dur": 5.204,
    "args": {
      "External id": 20,      "cbid": 41, "correlation": 502563
    }
  },
  {
    "ph": "s", "id": 502563, "pid": 14565, "tid": 14565, "ts": 1290507955547.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507955559.654, "dur": 0.515,
    "args": {
      "External id": 13,      "cbid": 317, "correlation": 502573
    }
  },
  {
    "ph": "f", "id": 502573, "pid": 14565, "tid": 14565, "ts": 1290507955559.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961005.588, "dur": 1.504,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961005.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961007.188, "dur": 1.312,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961007.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961008.596, "dur": 1.312,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961008.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961010.036, "dur": 1.120,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961010.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961011.444, "dur": 1.440,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961011.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961013.012, "dur": 1.152,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961013.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961014.260, "dur": 1.568,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961014.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961015.924, "dur": 1.152,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961015.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961017.204, "dur": 1.472,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961017.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961018.804, "dur": 1.152,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961018.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961020.052, "dur": 1.280,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961020.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961021.428, "dur": 1.216,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961021.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961022.740, "dur": 1.248,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961022.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961024.116, "dur": 1.408,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961024.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961025.652, "dur": 1.248,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961025.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507961027.028, "dur": 1.248,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961027.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507961028.436, "dur": 1.792,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961028.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507961030.388, "dur": 1.888,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961030.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507961032.436, "dur": 1.856,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961032.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507961034.420, "dur": 1.856,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961034.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961036.436, "dur": 2.496,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961036.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961039.028, "dur": 2.176,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961039.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961041.332, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961041.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961043.476, "dur": 2.144,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961043.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961045.748, "dur": 2.080,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961045.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961047.956, "dur": 1.888,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961047.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961049.972, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961049.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961052.116, "dur": 2.368,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961052.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961054.708, "dur": 1.920,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961054.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961056.724, "dur": 1.824,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961056.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961058.676, "dur": 1.600,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961058.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961060.596, "dur": 1.888,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961060.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961062.612, "dur": 1.664,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961062.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961064.372, "dur": 1.664,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961064.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961066.132, "dur": 1.984,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961066.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507961068.244, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961068.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961070.388, "dur": 3.520,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961070.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961074.036, "dur": 2.976,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961074.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961077.140, "dur": 2.144,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961077.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961079.412, "dur": 3.200,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961079.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961082.708, "dur": 2.560,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961082.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961085.364, "dur": 2.144,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961085.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961087.636, "dur": 2.496,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961087.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961090.260, "dur": 3.072,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961090.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961093.428, "dur": 2.784,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961093.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961096.340, "dur": 2.272,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961096.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961098.740, "dur": 3.200,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961098.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961102.068, "dur": 2.784,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961102.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961104.980, "dur": 2.656,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961104.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961107.732, "dur": 2.784,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961107.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961110.644, "dur": 2.784,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961110.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961113.556, "dur": 2.816,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961113.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961116.500, "dur": 2.432,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961116.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961119.060, "dur": 2.304,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961119.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961121.492, "dur": 2.464,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961121.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961124.084, "dur": 2.816,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961124.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961126.996, "dur": 2.720,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961126.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961129.812, "dur": 2.848,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961129.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961132.788, "dur": 2.560,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961132.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961135.476, "dur": 2.784,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961135.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961138.388, "dur": 2.400,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961138.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961140.916, "dur": 2.656,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961140.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961143.700, "dur": 2.368,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961143.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961146.196, "dur": 3.328,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961146.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961149.684, "dur": 2.688,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961149.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961152.532, "dur": 2.688,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961152.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961155.348, "dur": 3.072,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961155.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507961158.580, "dur": 2.592,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961158.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961161.300, "dur": 4.128,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961161.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961165.556, "dur": 4.800,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961165.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961170.484, "dur": 3.904,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961170.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961174.516, "dur": 4.992,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961174.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961179.604, "dur": 4.032,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961179.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961183.732, "dur": 4.384,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961183.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961188.244, "dur": 4.224,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961188.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961192.596, "dur": 4.480,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961192.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961197.172, "dur": 4.160,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961197.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961201.460, "dur": 4.224,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961201.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961205.812, "dur": 4.480,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961205.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961210.420, "dur": 4.352,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961210.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961214.900, "dur": 4.128,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961214.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961219.124, "dur": 5.088,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961219.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961224.308, "dur": 4.160,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961224.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507961228.596, "dur": 4.544,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961228.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290507961233.268, "dur": 7.328,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961233.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290507961240.724, "dur": 7.104,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961240.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961247.956, "dur": 12.224,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961247.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961260.308, "dur": 12.288,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961260.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961272.724, "dur": 12.128,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961272.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961284.948, "dur": 12.448,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961284.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290507961297.524, "dur": 15.872,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961297.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290507961313.524, "dur": 1.216,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961313.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961314.868, "dur": 2.944,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961314.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507961317.940, "dur": 21.696,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961317.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961339.764, "dur": 29.728,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961339.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507961369.620, "dur": 2.080,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961369.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507961371.796, "dur": 3.488,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961371.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507961375.412, "dur": 6.688,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961375.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507961382.196, "dur": 9.440,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961382.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507961391.732, "dur": 43.072,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961391.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507961434.932, "dur": 8.864,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961434.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961443.924, "dur": 17.216,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961443.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290507961461.300, "dur": 8.704,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961461.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961470.132, "dur": 48.896,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961470.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507961519.188, "dur": 13.440,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961519.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961532.788, "dur": 28.928,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961532.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290507961561.876, "dur": 22.496,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961561.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290507961584.500, "dur": 9.728,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961584.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507961593.204, "dur": 28.224,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961593.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507961621.556, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961621.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961623.668, "dur": 3.200,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961623.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961626.964, "dur": 29.760,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961626.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507961656.820, "dur": 2.048,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961656.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507961658.964, "dur": 3.424,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961658.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507961662.516, "dur": 6.464,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961662.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507961669.076, "dur": 8.928,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961669.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507961678.100, "dur": 40.864,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961678.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507961719.060, "dur": 8.832,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961719.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961728.020, "dur": 17.536,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961728.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507961745.684, "dur": 17.472,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961745.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961763.284, "dur": 51.296,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961763.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507961814.708, "dur": 13.312,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961814.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961828.180, "dur": 28.832,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961828.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507961857.140, "dur": 14.208,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961857.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507961871.476, "dur": 1.728,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961871.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961873.364, "dur": 3.808,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961873.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961877.268, "dur": 29.632,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961877.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507961907.028, "dur": 2.048,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961907.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507961909.204, "dur": 3.456,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961909.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507961912.788, "dur": 6.592,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961912.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507961919.476, "dur": 9.056,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961919.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507961928.628, "dur": 41.600,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961928.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507961970.356, "dur": 8.768,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961970.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507961979.252, "dur": 16.928,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961979.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507961996.308, "dur": 15.904,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507961996.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962012.341, "dur": 50.687,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962012.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507962063.156, "dur": 13.376,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962063.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962076.628, "dur": 28.704,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962076.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507962105.460, "dur": 23.744,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962105.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507962129.332, "dur": 1.953,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962129.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962131.380, "dur": 4.832,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962131.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962136.308, "dur": 30.208,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962136.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507962166.676, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962166.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507962168.820, "dur": 3.488,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962168.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507962172.436, "dur": 6.561,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962172.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507962179.124, "dur": 9.248,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962179.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507962188.501, "dur": 42.943,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962188.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507962231.604, "dur": 8.641,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962231.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962240.372, "dur": 15.969,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962240.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507962256.501, "dur": 13.055,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962256.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962269.652, "dur": 50.368,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962269.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507962320.149, "dur": 13.440,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962320.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962333.716, "dur": 27.809,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962333.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507962361.652, "dur": 14.496,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962361.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507962376.276, "dur": 1.792,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962376.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962378.196, "dur": 3.808,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962378.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962382.132, "dur": 29.472,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962382.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507962411.733, "dur": 2.047,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962411.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507962413.909, "dur": 3.487,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962413.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507962417.556, "dur": 6.497,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962417.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507962424.180, "dur": 9.024,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962424.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507962433.301, "dur": 41.791,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962433.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507962475.221, "dur": 8.511,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962475.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962483.860, "dur": 16.928,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962483.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507962500.916, "dur": 15.968,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962500.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962517.012, "dur": 50.880,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962517.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507962568.021, "dur": 13.247,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962568.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962581.396, "dur": 29.185,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962581.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507962610.741, "dur": 24.991,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962610.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507962634.740, "dur": 2.081,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962634.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962636.916, "dur": 4.352,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962636.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962641.364, "dur": 31.200,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962641.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507962672.693, "dur": 2.047,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962672.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507962674.869, "dur": 3.359,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962674.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507962678.388, "dur": 7.265,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962678.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507962685.780, "dur": 9.153,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962685.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507962695.028, "dur": 42.592,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962695.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507962737.717, "dur": 8.544,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962737.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962746.388, "dur": 17.409,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962746.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507962763.893, "dur": 12.831,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962763.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962776.853, "dur": 50.304,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962776.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507962827.317, "dur": 13.279,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962827.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962840.725, "dur": 28.800,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962840.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507962869.652, "dur": 14.336,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962869.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507962884.084, "dur": 1.793,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962884.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962886.004, "dur": 3.809,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962886.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962889.940, "dur": 29.696,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962889.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507962919.765, "dur": 2.047,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962919.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507962921.941, "dur": 3.456,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962921.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507962925.493, "dur": 6.623,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962925.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507962932.245, "dur": 9.184,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962932.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507962941.557, "dur": 42.432,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962941.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507962984.117, "dur": 8.767,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962984.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507962993.013, "dur": 16.640,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507962993.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507963009.780, "dur": 15.840,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963009.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963025.749, "dur": 50.464,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963025.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507963076.373, "dur": 13.375,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963076.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963089.909, "dur": 27.424,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963089.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507963117.461, "dur": 22.944,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963117.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507963140.501, "dur": 2.367,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963140.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963142.997, "dur": 4.928,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963142.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963148.053, "dur": 30.272,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963148.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507963178.421, "dur": 2.048,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963178.421,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507963180.597, "dur": 3.391,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963180.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507963184.117, "dur": 6.656,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963184.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507963190.901, "dur": 8.928,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963190.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507963199.924, "dur": 42.561,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963199.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507963242.613, "dur": 8.608,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963242.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963251.348, "dur": 17.665,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963251.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507963269.140, "dur": 12.833,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963269.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963282.133, "dur": 50.975,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963282.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507963333.237, "dur": 13.471,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963333.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963346.837, "dur": 28.032,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963346.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290507963374.964, "dur": 23.233,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963374.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290507963398.325, "dur": 8.608,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963398.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290507963407.061, "dur": 27.744,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963407.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507963434.933, "dur": 2.528,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963434.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963437.589, "dur": 3.392,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963437.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963441.141, "dur": 29.823,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963441.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507963471.093, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963471.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507963473.269, "dur": 3.392,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963473.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507963476.821, "dur": 6.688,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963476.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507963483.605, "dur": 9.152,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963483.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507963492.852, "dur": 43.456,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963492.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507963536.437, "dur": 8.896,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963536.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963545.461, "dur": 17.632,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963545.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507963563.188, "dur": 14.305,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963563.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963577.589, "dur": 51.584,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963577.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507963629.301, "dur": 13.312,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963629.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963642.709, "dur": 28.192,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963642.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507963671.029, "dur": 14.432,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963671.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507963685.589, "dur": 1.856,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963685.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963687.573, "dur": 4.096,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963687.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963691.797, "dur": 29.600,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963691.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507963721.525, "dur": 2.080,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963721.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507963723.701, "dur": 3.424,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963723.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507963727.253, "dur": 6.848,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963727.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507963734.229, "dur": 9.088,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963734.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507963743.381, "dur": 43.168,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963743.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507963786.677, "dur": 8.768,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963786.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963795.573, "dur": 17.024,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963795.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507963812.725, "dur": 15.968,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963812.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963828.821, "dur": 50.720,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963828.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507963879.669, "dur": 13.184,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963879.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963892.949, "dur": 28.736,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963892.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507963921.781, "dur": 21.120,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963921.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507963943.029, "dur": 1.600,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963943.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963944.757, "dur": 4.832,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963944.757,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507963949.717, "dur": 30.848,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963949.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507963980.693, "dur": 2.112,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963980.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507963982.965, "dur": 3.328,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963982.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507963986.453, "dur": 7.232,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963986.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507963993.813, "dur": 9.120,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507963993.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507964003.029, "dur": 41.984,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964003.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507964045.141, "dur": 8.416,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964045.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964053.653, "dur": 17.568,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964053.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507964071.349, "dur": 12.800,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964071.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964084.245, "dur": 51.424,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964084.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507964135.797, "dur": 13.408,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964135.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964149.301, "dur": 28.544,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964149.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507964177.973, "dur": 14.304,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964177.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507964192.405, "dur": 1.792,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964192.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964194.325, "dur": 4.224,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964194.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964198.677, "dur": 29.312,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964198.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507964228.117, "dur": 1.984,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964228.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507964230.229, "dur": 3.360,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964230.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507964233.717, "dur": 7.552,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964233.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507964241.365, "dur": 9.088,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964241.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507964250.581, "dur": 42.944,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964250.581,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507964293.653, "dur": 8.672,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964293.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964302.453, "dur": 17.088,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964302.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507964319.669, "dur": 15.936,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964319.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964335.733, "dur": 51.360,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964335.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507964387.221, "dur": 13.216,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964387.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964400.565, "dur": 28.416,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964400.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507964429.109, "dur": 22.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964429.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507964451.253, "dur": 1.920,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964451.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964453.301, "dur": 5.184,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964453.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964458.613, "dur": 30.048,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964458.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507964488.789, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964488.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507964490.933, "dur": 2.784,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964490.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507964493.845, "dur": 6.624,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964493.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507964500.565, "dur": 8.896,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964500.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507964509.557, "dur": 44.192,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964509.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507964553.877, "dur": 8.576,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964553.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964562.549, "dur": 17.280,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964562.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507964579.989, "dur": 13.120,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964579.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964593.237, "dur": 50.048,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964593.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507964643.413, "dur": 13.376,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964643.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964656.917, "dur": 29.248,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964656.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507964686.293, "dur": 14.560,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964686.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507964700.981, "dur": 1.888,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964700.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964702.997, "dur": 3.744,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964702.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964706.869, "dur": 29.728,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964706.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507964736.693, "dur": 2.080,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964736.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507964738.869, "dur": 3.456,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964738.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507964742.453, "dur": 6.624,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964742.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507964749.205, "dur": 8.992,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964749.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507964758.325, "dur": 41.984,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964758.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507964800.437, "dur": 8.768,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964800.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964809.333, "dur": 16.736,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964809.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507964826.165, "dur": 15.872,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964826.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964842.165, "dur": 50.592,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964842.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507964892.917, "dur": 13.376,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964892.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964906.421, "dur": 28.800,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964906.421,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507964935.317, "dur": 22.752,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964935.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507964958.165, "dur": 2.048,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964958.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964960.341, "dur": 5.440,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964960.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507964965.909, "dur": 29.632,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964965.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507964995.669, "dur": 2.080,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964995.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507964997.845, "dur": 3.616,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507964997.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507965001.589, "dur": 6.688,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965001.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507965008.405, "dur": 9.120,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965008.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507965017.621, "dur": 41.312,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965017.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507965059.061, "dur": 8.672,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965059.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965067.861, "dur": 17.408,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965067.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507965085.397, "dur": 12.832,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965085.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965098.357, "dur": 49.568,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965098.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507965148.085, "dur": 13.408,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965148.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965161.653, "dur": 28.032,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965161.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507965189.813, "dur": 14.400,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965189.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507965204.341, "dur": 1.856,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965204.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965206.325, "dur": 3.840,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965206.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965210.261, "dur": 29.376,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965210.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507965239.765, "dur": 2.016,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965239.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507965241.909, "dur": 3.328,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965241.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507965245.365, "dur": 6.656,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965245.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507965252.117, "dur": 8.928,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965252.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507965261.173, "dur": 43.968,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965261.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507965305.237, "dur": 8.576,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965305.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965313.973, "dur": 17.088,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965313.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507965331.189, "dur": 16.032,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965331.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965347.349, "dur": 50.240,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965347.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507965397.717, "dur": 13.376,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965397.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965411.189, "dur": 27.680,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965411.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290507965438.997, "dur": 12.992,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965438.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290507965452.117, "dur": 6.112,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965452.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965458.357, "dur": 148.192,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965458.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507965606.517, "dur": 145.216,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965606.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290507965751.637, "dur": 220.960,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965751.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290507965971.637, "dur": 20.448,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965971.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290507965991.989, "dur": 300.448,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507965991.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507966294.069, "dur": 0.512,
    "args": {
      "External id": 13,      "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "bytes": 192, "memory bandwidth (GB/s)": 0.375
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507966294.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507966295.573, "dur": 136.416,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507966295.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507966433.269, "dur": 1.088,
    "args": {
      "External id": 13,      "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "bytes": 576, "memory bandwidth (GB/s)": 0.5294117647058824
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507966433.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507966436.949, "dur": 140.160,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507966436.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290507966577.013, "dur": 185.952,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507966577.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290507966762.005, "dur": 20.320,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507966762.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290507966782.261, "dur": 302.816,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507966782.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507967086.485, "dur": 0.480,
    "args": {
      "External id": 13,      "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507967086.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967087.957, "dur": 138.368,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507967087.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507967228.949, "dur": 1.184,
    "args": {
      "External id": 13,      "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "bytes": 576, "memory bandwidth (GB/s)": 0.4864864864864865
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507967228.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967232.405, "dur": 140.512,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507967232.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290507967373.237, "dur": 3.328,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502575,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502575, "pid": 0, "tid": 7, "ts": 1290507967373.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507955560.894, "dur": 478.565,
    "args": {
      "External id": 13,      "cbid": 311, "correlation": 502575
    }
  },
  {
    "ph": "s", "id": 502575, "pid": 14565, "tid": 14565, "ts": 1290507955560.894,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290507967379.253, "dur": 1.568,
    "args": {
      "External id": 30,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502588,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502588, "pid": 0, "tid": 7, "ts": 1290507967379.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507957071.510, "dur": 19.169,
    "args": {
      "External id": 30,      "cbid": 211, "correlation": 502588
    }
  },
  {
    "ph": "s", "id": 502588, "pid": 14565, "tid": 14565, "ts": 1290507957071.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290507977840.727, "dur": 3.937,
    "args": {
      "External id": 38,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503411,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503411, "pid": 0, "tid": 7, "ts": 1290507977840.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507963104.113, "dur": 14.033,
    "args": {
      "External id": 38,      "cbid": 211, "correlation": 503411
    }
  },
  {
    "ph": "s", "id": 503411, "pid": 14565, "tid": 14565, "ts": 1290507963104.113,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507963125.366, "dur": 0.578,
    "args": {
      "External id": 36,      "cbid": 317, "correlation": 503419
    }
  },
  {
    "ph": "f", "id": 503419, "pid": 14565, "tid": 14565, "ts": 1290507963125.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290507977848.248, "dur": 0.832,
    "args": {
      "External id": 36,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503421,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503421, "pid": 0, "tid": 7, "ts": 1290507977848.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290507977849.207, "dur": 0.800,
    "args": {
      "External id": 36,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503421,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503421, "pid": 0, "tid": 7, "ts": 1290507977849.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290507977850.136, "dur": 2.080,
    "args": {
      "External id": 36,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503421,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503421, "pid": 0, "tid": 7, "ts": 1290507977850.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290507977852.343, "dur": 7.169,
    "args": {
      "External id": 36,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503421,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503421, "pid": 0, "tid": 7, "ts": 1290507977852.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290507977859.639, "dur": 10.689,
    "args": {
      "External id": 36,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503421,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503421, "pid": 0, "tid": 7, "ts": 1290507977859.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507963126.599, "dur": 14.384,
    "args": {
      "External id": 36,      "cbid": 311, "correlation": 503421
    }
  },
  {
    "ph": "s", "id": 503421, "pid": 14565, "tid": 14565, "ts": 1290507963126.599,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507977873.144, "dur": 1.376,
    "args": {
      "External id": 46,      "device": 0, "context": 1,      "stream": 7, "correlation": 503426,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 503426, "pid": 0, "tid": 7, "ts": 1290507977873.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507963567.761, "dur": 18.738,
    "args": {
      "External id": 46,      "cbid": 41, "correlation": 503426
    }
  },
  {
    "ph": "s", "id": 503426, "pid": 14565, "tid": 14565, "ts": 1290507963567.761,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507977874.904, "dur": 1.280,
    "args": {
      "External id": 47,      "device": 0, "context": 1,      "stream": 7, "correlation": 503431,      "bytes": 256, "memory bandwidth (GB/s)": 0.2
    }
  },
  {
    "ph": "f", "id": 503431, "pid": 0, "tid": 7, "ts": 1290507977874.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507963591.657, "dur": 4.960,
    "args": {
      "External id": 47,      "cbid": 41, "correlation": 503431
    }
  },
  {
    "ph": "s", "id": 503431, "pid": 14565, "tid": 14565, "ts": 1290507963591.657,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507977876.600, "dur": 1.376,
    "args": {
      "External id": 48,      "device": 0, "context": 1,      "stream": 7, "correlation": 503436,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 503436, "pid": 0, "tid": 7, "ts": 1290507977876.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507963599.643, "dur": 3.906,
    "args": {
      "External id": 48,      "cbid": 41, "correlation": 503436
    }
  },
  {
    "ph": "s", "id": 503436, "pid": 14565, "tid": 14565, "ts": 1290507963599.643,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507977878.424, "dur": 1.312,
    "args": {
      "External id": 49,      "device": 0, "context": 1,      "stream": 7, "correlation": 503441,      "bytes": 256, "memory bandwidth (GB/s)": 0.1951219512195122
    }
  },
  {
    "ph": "f", "id": 503441, "pid": 0, "tid": 7, "ts": 1290507977878.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507963606.451, "dur": 4.640,
    "args": {
      "External id": 49,      "cbid": 41, "correlation": 503441
    }
  },
  {
    "ph": "s", "id": 503441, "pid": 14565, "tid": 14565, "ts": 1290507963606.451,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507977880.183, "dur": 2.145,
    "args": {
      "External id": 50,      "device": 0, "context": 1,      "stream": 7, "correlation": 503446,      "bytes": 65536, "memory bandwidth (GB/s)": 30.552913752913753
    }
  },
  {
    "ph": "f", "id": 503446, "pid": 0, "tid": 7, "ts": 1290507977880.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507963613.960, "dur": 3.953,
    "args": {
      "External id": 50,      "cbid": 41, "correlation": 503446
    }
  },
  {
    "ph": "s", "id": 503446, "pid": 14565, "tid": 14565, "ts": 1290507963613.960,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507963624.698, "dur": 0.531,
    "args": {
      "External id": 43,      "cbid": 317, "correlation": 503456
    }
  },
  {
    "ph": "f", "id": 503456, "pid": 14565, "tid": 14565, "ts": 1290507963624.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977887.255, "dur": 1.473,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977887.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977888.855, "dur": 1.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977888.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977890.231, "dur": 1.313,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977890.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977891.640, "dur": 1.247,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977891.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977893.143, "dur": 1.440,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977893.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977894.712, "dur": 1.184,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977894.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977896.088, "dur": 1.471,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977896.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977897.656, "dur": 1.152,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977897.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977898.935, "dur": 1.473,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977898.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977900.504, "dur": 1.184,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977900.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977901.784, "dur": 1.344,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977901.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977903.224, "dur": 1.312,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977903.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977904.664, "dur": 1.247,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977904.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977906.007, "dur": 1.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977906.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977907.416, "dur": 1.216,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977907.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507977908.760, "dur": 1.248,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977908.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507977910.168, "dur": 2.271,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977910.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507977912.599, "dur": 1.856,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977912.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507977914.744, "dur": 2.111,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977914.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507977916.984, "dur": 1.823,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977916.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977918.904, "dur": 2.079,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977918.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977921.207, "dur": 2.209,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977921.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977923.543, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977923.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977925.688, "dur": 2.080,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977925.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977927.895, "dur": 1.825,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977927.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977929.815, "dur": 2.081,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977929.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977932.024, "dur": 1.983,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977932.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977934.136, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977934.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977936.439, "dur": 1.920,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977936.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977938.488, "dur": 2.207,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977938.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977940.792, "dur": 1.792,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977940.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977942.871, "dur": 1.696,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977942.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977944.696, "dur": 1.728,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977944.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977946.551, "dur": 1.825,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977946.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977948.503, "dur": 2.336,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977948.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507977950.968, "dur": 1.984,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977950.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977953.079, "dur": 3.136,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977953.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977956.311, "dur": 3.265,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977956.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977959.704, "dur": 2.368,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977959.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977962.167, "dur": 2.785,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977962.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977965.079, "dur": 2.625,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977965.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977967.832, "dur": 2.112,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977967.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977970.071, "dur": 2.496,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977970.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977972.696, "dur": 2.784,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977972.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977975.607, "dur": 3.136,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977975.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977978.872, "dur": 2.400,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977978.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977981.399, "dur": 2.752,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977981.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977984.280, "dur": 2.784,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977984.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977987.191, "dur": 2.912,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977987.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977990.232, "dur": 2.527,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977990.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977992.888, "dur": 2.751,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977992.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977995.768, "dur": 2.976,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977995.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507977998.871, "dur": 2.400,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507977998.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978001.367, "dur": 2.496,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978001.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978003.960, "dur": 3.104,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978003.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978007.192, "dur": 2.752,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978007.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978010.071, "dur": 2.337,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978010.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978012.504, "dur": 2.783,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978012.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978015.383, "dur": 2.656,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978015.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978018.168, "dur": 2.816,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978018.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978021.079, "dur": 2.496,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978021.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978023.671, "dur": 2.945,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978023.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978026.744, "dur": 2.303,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978026.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978029.176, "dur": 3.136,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978029.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978032.439, "dur": 3.296,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978032.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978035.864, "dur": 2.751,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978035.864,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978038.775, "dur": 2.625,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978038.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507978041.560, "dur": 2.912,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978041.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978044.600, "dur": 4.255,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978044.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978048.984, "dur": 4.703,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978048.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978053.816, "dur": 3.807,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978053.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978057.720, "dur": 4.639,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978057.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978062.455, "dur": 4.033,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978062.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978066.616, "dur": 4.031,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978066.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978070.776, "dur": 4.512,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978070.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978075.415, "dur": 4.033,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978075.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978079.544, "dur": 4.576,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978079.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978084.216, "dur": 3.936,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978084.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978088.279, "dur": 5.025,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978088.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978093.431, "dur": 4.001,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978093.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978097.560, "dur": 4.799,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978097.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978102.488, "dur": 4.416,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978102.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978107.000, "dur": 4.416,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978107.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507978111.512, "dur": 4.416,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978111.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290507978116.056, "dur": 7.616,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978116.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290507978123.800, "dur": 7.264,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978123.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978131.192, "dur": 12.607,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978131.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978143.928, "dur": 12.480,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978143.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978156.536, "dur": 12.287,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978156.536,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978168.952, "dur": 12.448,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978168.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290507978181.560, "dur": 16.032,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978181.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290507978197.719, "dur": 1.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978197.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978199.128, "dur": 2.976,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978199.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507978202.264, "dur": 21.888,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978202.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978224.279, "dur": 29.953,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978224.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507978254.360, "dur": 2.111,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978254.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507978256.568, "dur": 3.519,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978256.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507978260.184, "dur": 7.488,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978260.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507978267.800, "dur": 9.376,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978267.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507978277.272, "dur": 41.760,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978277.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507978319.159, "dur": 8.705,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978319.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978327.960, "dur": 17.728,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978327.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290507978345.848, "dur": 8.608,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978345.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978354.584, "dur": 49.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978354.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507978403.992, "dur": 13.503,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978403.992,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978417.624, "dur": 29.312,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978417.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290507978447.096, "dur": 21.983,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978447.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290507978469.208, "dur": 9.376,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978469.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507978477.752, "dur": 28.640,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978477.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507978506.520, "dur": 1.504,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978506.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978508.120, "dur": 3.456,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978508.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978511.672, "dur": 29.760,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978511.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507978541.528, "dur": 2.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978541.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507978543.704, "dur": 3.328,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978543.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507978547.160, "dur": 6.720,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978547.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507978554.008, "dur": 9.024,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978554.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507978563.128, "dur": 44.960,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978563.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507978608.216, "dur": 9.088,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978608.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978617.464, "dur": 17.696,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978617.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507978635.320, "dur": 16.160,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978635.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978651.608, "dur": 51.264,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978651.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507978703.032, "dur": 13.248,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978703.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978716.408, "dur": 28.864,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978716.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507978745.368, "dur": 14.144,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978745.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507978759.640, "dur": 1.888,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978759.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978761.656, "dur": 3.936,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978761.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978765.688, "dur": 29.408,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978765.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507978795.192, "dur": 2.080,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978795.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507978797.400, "dur": 3.456,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978797.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507978800.984, "dur": 6.816,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978800.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507978807.896, "dur": 9.344,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978807.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507978817.368, "dur": 43.360,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978817.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507978860.856, "dur": 8.960,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978860.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978869.912, "dur": 16.768,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978869.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507978886.808, "dur": 15.872,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978886.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978902.808, "dur": 51.072,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978902.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507978953.976, "dur": 13.312,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978953.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507978967.384, "dur": 29.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978967.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507978996.792, "dur": 25.152,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507978996.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507979020.984, "dur": 3.168,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979020.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979023.128, "dur": 5.408,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979023.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979028.600, "dur": 29.728,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979028.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507979058.456, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979058.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507979060.632, "dur": 3.520,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979060.632,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507979064.280, "dur": 6.592,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979064.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507979071.000, "dur": 9.056,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979071.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507979080.184, "dur": 42.560,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979080.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507979122.872, "dur": 8.608,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979122.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979131.608, "dur": 16.192,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979131.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507979147.960, "dur": 13.184,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979147.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979161.240, "dur": 50.912,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979161.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507979212.248, "dur": 13.568,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979212.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979225.944, "dur": 28.480,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979225.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507979254.552, "dur": 14.496,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979254.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507979269.144, "dur": 1.792,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979269.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979271.064, "dur": 3.840,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979271.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979275.032, "dur": 29.792,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979275.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507979304.952, "dur": 2.080,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979304.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507979307.160, "dur": 3.456,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979307.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507979310.744, "dur": 6.816,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979310.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507979317.656, "dur": 9.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979317.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507979327.032, "dur": 42.336,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979327.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507979369.496, "dur": 8.640,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979369.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979378.264, "dur": 17.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979378.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507979395.672, "dur": 15.936,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979395.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979411.736, "dur": 50.432,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979411.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507979462.296, "dur": 13.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979462.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979475.672, "dur": 30.432,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979475.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507979506.264, "dur": 24.704,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979506.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507979529.848, "dur": 1.984,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979529.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979531.960, "dur": 4.320,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979531.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979536.408, "dur": 31.168,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979536.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507979567.704, "dur": 2.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979567.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507979569.848, "dur": 3.392,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979569.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507979573.368, "dur": 6.592,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979573.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507979580.088, "dur": 8.928,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979580.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507979589.144, "dur": 41.664,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979589.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507979630.936, "dur": 8.384,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979630.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979639.448, "dur": 17.344,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979639.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507979656.920, "dur": 12.928,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979656.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979669.976, "dur": 50.112,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979669.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507979720.248, "dur": 13.344,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979720.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979733.720, "dur": 28.960,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979733.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507979762.808, "dur": 14.464,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979762.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507979777.368, "dur": 1.696,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979777.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979779.192, "dur": 3.968,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979779.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979783.256, "dur": 29.632,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979783.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507979813.016, "dur": 2.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979813.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507979815.160, "dur": 3.488,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979815.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507979818.776, "dur": 7.392,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979818.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507979826.296, "dur": 9.024,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979826.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507979835.448, "dur": 42.400,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979835.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507979877.976, "dur": 8.640,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979877.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979886.744, "dur": 16.928,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979886.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507979903.800, "dur": 15.840,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979903.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979919.768, "dur": 50.688,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979919.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507979970.616, "dur": 13.408,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979970.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507979984.184, "dur": 27.552,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507979984.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507980011.832, "dur": 23.040,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980011.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507980034.968, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980034.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980037.080, "dur": 4.704,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980037.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980041.912, "dur": 30.304,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980041.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507980072.344, "dur": 2.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980072.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507980074.488, "dur": 4.224,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980074.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507980078.840, "dur": 6.624,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980078.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507980085.560, "dur": 8.960,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980085.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507980094.616, "dur": 43.936,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980094.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507980138.680, "dur": 8.576,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980138.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980147.384, "dur": 17.952,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980147.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507980165.496, "dur": 12.736,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980165.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980178.360, "dur": 51.136,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980178.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507980229.656, "dur": 13.376,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980229.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980243.160, "dur": 28.672,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980243.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290507980271.960, "dur": 22.336,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980271.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290507980294.424, "dur": 8.608,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980294.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290507980303.160, "dur": 27.776,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980303.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507980331.064, "dur": 2.176,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980331.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980333.368, "dur": 3.360,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980333.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980336.856, "dur": 29.536,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980336.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507980366.520, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980366.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507980368.664, "dur": 3.488,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980368.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507980372.280, "dur": 6.784,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980372.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507980379.160, "dur": 9.216,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980379.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507980388.472, "dur": 41.792,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980388.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507980430.392, "dur": 8.672,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980430.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980439.192, "dur": 17.664,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980439.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507980456.984, "dur": 15.968,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980456.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980473.080, "dur": 50.368,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980473.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507980523.544, "dur": 13.216,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980523.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980536.888, "dur": 27.904,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980536.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507980564.952, "dur": 14.208,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980564.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507980579.288, "dur": 1.664,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980579.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980581.080, "dur": 4.000,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980581.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980585.208, "dur": 29.376,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980585.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507980614.712, "dur": 2.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980614.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507980616.888, "dur": 3.424,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980616.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507980620.440, "dur": 6.720,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980620.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507980627.288, "dur": 9.088,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980627.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507980636.504, "dur": 41.440,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980636.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507980678.072, "dur": 8.544,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980678.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980686.776, "dur": 17.280,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980686.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507980704.184, "dur": 15.904,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980704.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980720.216, "dur": 50.656,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980720.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507980771.000, "dur": 13.344,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980771.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980784.472, "dur": 28.544,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980784.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507980813.144, "dur": 21.600,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980813.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507980834.840, "dur": 2.144,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980834.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980837.080, "dur": 4.768,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980837.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980841.976, "dur": 30.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980841.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507980872.120, "dur": 1.984,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980872.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507980874.264, "dur": 3.520,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980874.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507980877.912, "dur": 6.656,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980877.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507980884.696, "dur": 8.864,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980884.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507980893.688, "dur": 43.328,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980893.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507980937.144, "dur": 8.288,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980937.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980945.560, "dur": 17.920,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980945.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507980963.608, "dur": 12.736,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980963.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507980976.472, "dur": 49.664,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507980976.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507981026.264, "dur": 13.440,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981026.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981039.832, "dur": 29.312,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981039.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507981069.272, "dur": 14.304,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981069.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507981083.704, "dur": 1.760,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981083.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981085.592, "dur": 4.064,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981085.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981089.784, "dur": 29.312,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981089.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507981119.224, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981119.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507981121.368, "dur": 3.360,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981121.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507981124.856, "dur": 6.784,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981124.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507981131.768, "dur": 9.088,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981131.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507981140.984, "dur": 41.696,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981140.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507981182.808, "dur": 8.576,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981182.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981191.544, "dur": 17.024,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981191.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507981208.696, "dur": 15.872,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981208.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981224.696, "dur": 50.880,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981224.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507981275.704, "dur": 13.312,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981275.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981289.144, "dur": 28.896,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981289.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507981318.168, "dur": 21.504,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981318.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507981339.800, "dur": 2.112,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981339.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981342.040, "dur": 5.856,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981342.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981348.024, "dur": 29.856,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981348.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507981378.008, "dur": 2.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981378.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507981380.184, "dur": 3.360,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981380.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507981383.672, "dur": 6.560,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981383.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507981390.328, "dur": 8.960,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981390.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507981399.416, "dur": 40.416,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981399.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507981439.960, "dur": 8.288,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981439.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981448.376, "dur": 17.536,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981448.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507981466.072, "dur": 13.152,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981466.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981479.384, "dur": 50.208,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981479.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507981529.720, "dur": 13.440,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981529.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981543.320, "dur": 29.536,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981543.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507981572.984, "dur": 14.464,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981572.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507981587.576, "dur": 1.824,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981587.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981589.560, "dur": 4.128,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981589.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981593.848, "dur": 29.792,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981593.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507981623.736, "dur": 2.080,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981623.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507981625.912, "dur": 3.392,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981625.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507981629.432, "dur": 6.720,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981629.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507981636.248, "dur": 9.024,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981636.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507981645.400, "dur": 44.736,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981645.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507981690.264, "dur": 8.896,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981690.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981699.288, "dur": 16.896,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981699.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507981716.312, "dur": 15.968,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981716.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981732.408, "dur": 50.528,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981732.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507981783.096, "dur": 13.408,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981783.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981796.664, "dur": 28.320,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981796.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507981825.112, "dur": 23.040,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981825.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507981848.280, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981848.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981850.392, "dur": 4.704,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981850.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981855.256, "dur": 30.336,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981855.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507981885.720, "dur": 2.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981885.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507981887.896, "dur": 3.456,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981887.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507981891.480, "dur": 6.560,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981891.480,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507981898.136, "dur": 8.960,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981898.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507981907.192, "dur": 42.144,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981907.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507981949.464, "dur": 8.640,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981949.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981958.232, "dur": 17.153,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981958.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507981975.512, "dur": 12.800,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981975.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507981988.440, "dur": 49.856,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507981988.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507982038.456, "dur": 13.472,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982038.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982052.056, "dur": 27.776,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982052.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507982079.960, "dur": 14.240,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982079.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507982094.329, "dur": 1.791,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982094.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982096.216, "dur": 4.000,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982096.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982100.344, "dur": 29.696,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982100.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507982130.168, "dur": 2.016,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982130.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507982132.313, "dur": 4.256,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982132.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507982136.696, "dur": 7.456,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982136.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507982144.280, "dur": 8.960,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982144.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507982153.336, "dur": 42.112,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982153.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507982195.608, "dur": 8.481,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982195.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982204.249, "dur": 16.863,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982204.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507982221.240, "dur": 16.224,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982221.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982237.592, "dur": 50.113,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982237.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507982287.832, "dur": 13.376,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982287.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982301.336, "dur": 27.776,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982301.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290507982329.176, "dur": 12.992,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982329.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290507982342.296, "dur": 6.336,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982342.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982348.728, "dur": 147.329,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982348.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507982495.992, "dur": 144.033,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982495.992,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290507982639.960, "dur": 222.401,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982639.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290507982861.145, "dur": 20.768,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982861.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290507982881.849, "dur": 300.544,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507982881.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507983183.801, "dur": 0.480,
    "args": {
      "External id": 43,      "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983183.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507983185.305, "dur": 137.984,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983185.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507983325.081, "dur": 1.152,
    "args": {
      "External id": 43,      "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "bytes": 576, "memory bandwidth (GB/s)": 0.5
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983325.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507983328.921, "dur": 141.600,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983328.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290507983470.425, "dur": 186.688,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983470.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290507983656.089, "dur": 20.768,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983656.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290507983676.825, "dur": 298.048,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983676.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507983977.369, "dur": 0.480,
    "args": {
      "External id": 43,      "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983977.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507983978.873, "dur": 138.912,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507983978.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507984120.857, "dur": 1.184,
    "args": {
      "External id": 43,      "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "bytes": 576, "memory bandwidth (GB/s)": 0.4864864864864865
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507984120.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984123.673, "dur": 141.248,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507984123.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290507984265.241, "dur": 3.360,
    "args": {
      "External id": 43,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503458,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503458, "pid": 0, "tid": 7, "ts": 1290507984265.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507963625.820, "dur": 385.756,
    "args": {
      "External id": 43,      "cbid": 311, "correlation": 503458
    }
  },
  {
    "ph": "s", "id": 503458, "pid": 14565, "tid": 14565, "ts": 1290507963625.820,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290507984271.161, "dur": 1.504,
    "args": {
      "External id": 60,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503471,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503471, "pid": 0, "tid": 7, "ts": 1290507984271.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507964757.606, "dur": 12.037,
    "args": {
      "External id": 60,      "cbid": 211, "correlation": 503471
    }
  },
  {
    "ph": "s", "id": 503471, "pid": 14565, "tid": 14565, "ts": 1290507964757.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290507994728.475, "dur": 3.872,
    "args": {
      "External id": 68,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504294,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504294, "pid": 0, "tid": 7, "ts": 1290507994728.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507971629.830, "dur": 16.128,
    "args": {
      "External id": 68,      "cbid": 211, "correlation": 504294
    }
  },
  {
    "ph": "s", "id": 504294, "pid": 14565, "tid": 14565, "ts": 1290507971629.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507971653.621, "dur": 0.584,
    "args": {
      "External id": 66,      "cbid": 317, "correlation": 504302
    }
  },
  {
    "ph": "f", "id": 504302, "pid": 14565, "tid": 14565, "ts": 1290507971653.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290507994736.155, "dur": 0.832,
    "args": {
      "External id": 66,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504304,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504304, "pid": 0, "tid": 7, "ts": 1290507994736.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290507994737.147, "dur": 0.800,
    "args": {
      "External id": 66,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504304,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504304, "pid": 0, "tid": 7, "ts": 1290507994737.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290507994738.075, "dur": 2.176,
    "args": {
      "External id": 66,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504304,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504304, "pid": 0, "tid": 7, "ts": 1290507994738.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290507994740.411, "dur": 6.976,
    "args": {
      "External id": 66,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504304,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504304, "pid": 0, "tid": 7, "ts": 1290507994740.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290507994747.515, "dur": 10.624,
    "args": {
      "External id": 66,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504304,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504304, "pid": 0, "tid": 7, "ts": 1290507994747.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507971654.862, "dur": 20.855,
    "args": {
      "External id": 66,      "cbid": 311, "correlation": 504304
    }
  },
  {
    "ph": "s", "id": 504304, "pid": 14565, "tid": 14565, "ts": 1290507971654.862,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507994760.955, "dur": 1.376,
    "args": {
      "External id": 76,      "device": 0, "context": 1,      "stream": 7, "correlation": 504309,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 504309, "pid": 0, "tid": 7, "ts": 1290507994760.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507972352.141, "dur": 25.675,
    "args": {
      "External id": 76,      "cbid": 41, "correlation": 504309
    }
  },
  {
    "ph": "s", "id": 504309, "pid": 14565, "tid": 14565, "ts": 1290507972352.141,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507994762.747, "dur": 1.248,
    "args": {
      "External id": 77,      "device": 0, "context": 1,      "stream": 7, "correlation": 504314,      "bytes": 256, "memory bandwidth (GB/s)": 0.20512820512820512
    }
  },
  {
    "ph": "f", "id": 504314, "pid": 0, "tid": 7, "ts": 1290507994762.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507972385.594, "dur": 6.221,
    "args": {
      "External id": 77,      "cbid": 41, "correlation": 504314
    }
  },
  {
    "ph": "s", "id": 504314, "pid": 14565, "tid": 14565, "ts": 1290507972385.594,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507994764.443, "dur": 1.376,
    "args": {
      "External id": 78,      "device": 0, "context": 1,      "stream": 7, "correlation": 504319,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 504319, "pid": 0, "tid": 7, "ts": 1290507994764.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507972395.402, "dur": 4.797,
    "args": {
      "External id": 78,      "cbid": 41, "correlation": 504319
    }
  },
  {
    "ph": "s", "id": 504319, "pid": 14565, "tid": 14565, "ts": 1290507972395.402,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507994766.267, "dur": 1.312,
    "args": {
      "External id": 79,      "device": 0, "context": 1,      "stream": 7, "correlation": 504324,      "bytes": 256, "memory bandwidth (GB/s)": 0.1951219512195122
    }
  },
  {
    "ph": "f", "id": 504324, "pid": 0, "tid": 7, "ts": 1290507994766.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507972403.722, "dur": 5.629,
    "args": {
      "External id": 79,      "cbid": 41, "correlation": 504324
    }
  },
  {
    "ph": "s", "id": 504324, "pid": 14565, "tid": 14565, "ts": 1290507972403.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290507994768.027, "dur": 2.016,
    "args": {
      "External id": 80,      "device": 0, "context": 1,      "stream": 7, "correlation": 504329,      "bytes": 65536, "memory bandwidth (GB/s)": 32.507936507936506
    }
  },
  {
    "ph": "f", "id": 504329, "pid": 0, "tid": 7, "ts": 1290507994768.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507972413.452, "dur": 5.185,
    "args": {
      "External id": 80,      "cbid": 41, "correlation": 504329
    }
  },
  {
    "ph": "s", "id": 504329, "pid": 14565, "tid": 14565, "ts": 1290507972413.452,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507972427.543, "dur": 0.569,
    "args": {
      "External id": 73,      "cbid": 317, "correlation": 504339
    }
  },
  {
    "ph": "f", "id": 504339, "pid": 14565, "tid": 14565, "ts": 1290507972427.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994775.067, "dur": 1.504,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994775.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994776.667, "dur": 1.312,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994776.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994778.075, "dur": 1.312,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994778.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994779.483, "dur": 1.152,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994779.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994780.923, "dur": 1.440,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994780.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994782.491, "dur": 1.152,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994782.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994783.771, "dur": 1.696,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994783.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994785.563, "dur": 1.184,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994785.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994786.875, "dur": 1.472,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994786.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994788.475, "dur": 1.152,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994788.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994789.723, "dur": 1.504,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994789.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994791.323, "dur": 1.216,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994791.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994792.667, "dur": 1.344,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994792.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994794.139, "dur": 1.280,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994794.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994795.547, "dur": 1.248,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994795.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290507994796.923, "dur": 1.248,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994796.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507994798.331, "dur": 1.824,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994798.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507994800.315, "dur": 1.792,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994800.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507994802.267, "dur": 2.080,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994802.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290507994804.507, "dur": 1.824,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994804.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994806.459, "dur": 2.624,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994806.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994809.211, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994809.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994811.323, "dur": 2.080,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994811.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994813.499, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994813.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994815.611, "dur": 2.208,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994815.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994818.043, "dur": 2.272,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994818.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994820.411, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994820.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994822.587, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994822.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994824.731, "dur": 1.696,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994824.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994826.555, "dur": 1.824,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994826.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994828.475, "dur": 1.632,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994828.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994830.427, "dur": 1.888,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994830.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994832.443, "dur": 1.664,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994832.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994834.235, "dur": 1.632,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994834.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994835.963, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994835.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290507994838.107, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994838.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994840.251, "dur": 3.328,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994840.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994843.707, "dur": 3.360,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994843.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994847.195, "dur": 2.240,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994847.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994849.563, "dur": 2.720,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994849.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994852.411, "dur": 2.880,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994852.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994855.419, "dur": 2.176,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994855.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994857.691, "dur": 2.592,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994857.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994860.411, "dur": 2.912,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994860.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994863.451, "dur": 3.072,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994863.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994866.651, "dur": 2.560,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994866.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994869.339, "dur": 2.944,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994869.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994872.379, "dur": 3.104,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994872.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994875.611, "dur": 2.624,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994875.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994878.363, "dur": 2.624,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994878.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994881.115, "dur": 3.200,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994881.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994884.443, "dur": 2.912,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994884.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994887.451, "dur": 2.304,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994887.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994889.851, "dur": 2.464,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994889.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994892.443, "dur": 2.592,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994892.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994895.163, "dur": 2.752,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994895.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994898.043, "dur": 2.208,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994898.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994900.379, "dur": 3.168,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994900.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994903.675, "dur": 2.528,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994903.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994906.331, "dur": 2.656,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994906.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994909.083, "dur": 2.496,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994909.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994911.707, "dur": 2.560,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994911.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994914.395, "dur": 2.528,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994914.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994917.051, "dur": 3.104,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994917.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994920.315, "dur": 2.848,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994920.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994923.323, "dur": 2.816,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994923.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994926.267, "dur": 2.528,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994926.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290507994928.955, "dur": 2.944,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994928.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994932.027, "dur": 4.256,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994932.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994936.411, "dur": 4.864,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994936.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994941.371, "dur": 3.808,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994941.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994945.307, "dur": 5.120,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994945.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994950.523, "dur": 4.160,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994950.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994954.779, "dur": 4.096,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994954.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994958.971, "dur": 4.128,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994958.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994963.227, "dur": 4.224,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994963.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994967.547, "dur": 4.448,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994967.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994972.091, "dur": 3.872,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994972.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994976.091, "dur": 4.960,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994976.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994981.211, "dur": 4.000,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994981.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994985.307, "dur": 4.448,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994985.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994989.883, "dur": 4.576,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994989.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994994.587, "dur": 4.192,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994994.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290507994998.907, "dur": 4.480,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507994998.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290507995003.515, "dur": 7.424,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995003.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290507995011.067, "dur": 7.200,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995011.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995018.395, "dur": 12.512,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995018.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995031.035, "dur": 12.608,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995031.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995043.739, "dur": 12.352,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995043.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995056.219, "dur": 12.672,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995056.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290507995069.051, "dur": 16.064,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995069.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290507995085.275, "dur": 1.248,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995085.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995086.651, "dur": 2.976,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995086.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507995089.755, "dur": 22.112,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995089.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995111.963, "dur": 29.856,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995111.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507995141.915, "dur": 2.112,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995141.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507995144.123, "dur": 3.520,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995144.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507995147.771, "dur": 6.720,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995147.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507995154.587, "dur": 9.376,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995154.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507995164.059, "dur": 44.384,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995164.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507995208.571, "dur": 8.928,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995208.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995217.627, "dur": 17.856,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995217.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290507995235.611, "dur": 8.608,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995235.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995244.379, "dur": 49.152,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995244.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507995293.691, "dur": 13.504,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995293.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995307.323, "dur": 29.504,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995307.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290507995336.955, "dur": 22.080,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995336.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290507995359.195, "dur": 9.536,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995359.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507995367.675, "dur": 29.216,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995367.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507995397.019, "dur": 1.888,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995397.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995399.003, "dur": 3.296,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995399.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995402.395, "dur": 29.888,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995402.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507995432.411, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995432.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507995434.555, "dur": 3.424,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995434.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507995438.107, "dur": 7.232,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995438.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507995445.467, "dur": 9.120,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995445.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507995454.683, "dur": 41.952,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995454.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507995496.731, "dur": 8.928,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995496.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995505.819, "dur": 17.696,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995505.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507995523.643, "dur": 17.248,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995523.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995541.019, "dur": 51.488,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995541.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507995592.667, "dur": 13.472,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995592.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995606.267, "dur": 29.664,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995606.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507995636.059, "dur": 14.112,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995636.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507995650.299, "dur": 1.792,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995650.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995652.219, "dur": 3.904,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995652.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995656.187, "dur": 29.920,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995656.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507995686.235, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995686.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507995688.443, "dur": 3.424,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995688.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507995691.995, "dur": 7.648,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995691.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507995699.739, "dur": 9.376,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995699.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507995709.211, "dur": 42.944,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995709.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507995752.251, "dur": 8.864,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995752.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995761.243, "dur": 17.344,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995761.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507995778.715, "dur": 15.872,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995778.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995794.715, "dur": 50.944,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995794.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507995845.787, "dur": 13.376,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995845.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995859.259, "dur": 29.216,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995859.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507995888.603, "dur": 23.136,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995888.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507995911.835, "dur": 3.136,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995911.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995915.067, "dur": 4.928,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995915.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507995920.091, "dur": 29.824,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995920.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507995950.075, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995950.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507995952.219, "dur": 3.520,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995952.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507995955.899, "dur": 7.488,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995955.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507995963.483, "dur": 9.120,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995963.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507995972.731, "dur": 42.304,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507995972.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507996015.163, "dur": 8.544,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996015.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996023.835, "dur": 16.224,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996023.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507996040.219, "dur": 13.056,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996040.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996053.371, "dur": 49.728,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996053.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507996103.195, "dur": 13.568,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996103.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996116.891, "dur": 28.128,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996116.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507996145.147, "dur": 14.400,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996145.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507996159.675, "dur": 1.760,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996159.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996161.563, "dur": 3.808,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996161.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996165.499, "dur": 29.632,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996165.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507996195.259, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996195.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507996197.435, "dur": 3.456,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996197.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507996201.019, "dur": 6.592,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996201.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507996207.707, "dur": 9.056,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996207.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507996216.891, "dur": 42.240,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996216.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507996259.259, "dur": 8.576,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996259.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996267.963, "dur": 16.928,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996267.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507996285.019, "dur": 15.968,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996285.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996301.115, "dur": 50.432,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996301.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507996351.675, "dur": 13.440,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996351.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996365.211, "dur": 29.568,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996365.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507996394.907, "dur": 23.936,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996394.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507996417.819, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996417.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996419.995, "dur": 4.288,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996419.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996424.411, "dur": 31.008,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996424.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507996455.547, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996455.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507996457.659, "dur": 4.128,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996457.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507996461.915, "dur": 6.560,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996461.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507996468.603, "dur": 8.960,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996468.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507996477.659, "dur": 42.272,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996477.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507996520.059, "dur": 8.480,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996520.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996528.667, "dur": 17.664,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996528.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507996546.427, "dur": 12.864,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996546.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996559.419, "dur": 50.336,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996559.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507996609.915, "dur": 13.408,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996609.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996623.451, "dur": 28.832,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996623.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507996652.411, "dur": 14.240,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996652.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507996666.779, "dur": 1.888,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996666.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996668.795, "dur": 4.000,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996668.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996672.923, "dur": 29.984,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996672.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507996703.035, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996703.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507996705.179, "dur": 3.680,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996705.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507996708.955, "dur": 6.688,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996708.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507996715.771, "dur": 9.088,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996715.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507996724.955, "dur": 42.688,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996724.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507996767.739, "dur": 8.672,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996767.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996776.539, "dur": 16.832,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996776.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507996793.531, "dur": 15.872,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996793.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996809.499, "dur": 51.264,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996809.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507996860.923, "dur": 13.376,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996860.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996874.459, "dur": 28.064,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996874.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507996902.651, "dur": 22.944,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996902.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507996925.691, "dur": 2.144,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996925.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996927.931, "dur": 4.896,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996927.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507996932.923, "dur": 30.912,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996932.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507996963.963, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996963.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507996966.107, "dur": 3.296,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996966.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507996969.531, "dur": 6.848,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996969.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507996976.507, "dur": 9.120,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996976.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507996985.723, "dur": 41.760,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507996985.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507997027.611, "dur": 8.512,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997027.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997036.251, "dur": 17.792,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997036.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507997054.203, "dur": 12.864,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997054.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997067.195, "dur": 50.848,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997067.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507997118.203, "dur": 13.472,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997118.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997131.803, "dur": 28.544,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997131.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290507997160.475, "dur": 22.560,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997160.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290507997183.132, "dur": 8.511,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997183.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290507997191.803, "dur": 28.224,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997191.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507997220.156, "dur": 1.471,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997220.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997221.756, "dur": 3.231,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997221.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997225.115, "dur": 29.984,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997225.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507997255.227, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997255.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507997257.404, "dur": 3.295,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997257.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507997260.827, "dur": 7.104,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997260.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507997268.028, "dur": 9.119,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997268.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507997277.243, "dur": 43.936,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997277.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507997321.308, "dur": 8.831,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997321.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997330.267, "dur": 17.728,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997330.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507997348.123, "dur": 14.400,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997348.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997362.619, "dur": 52.033,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997362.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507997414.779, "dur": 13.248,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997414.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997428.155, "dur": 27.968,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997428.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507997456.252, "dur": 14.463,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997456.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507997470.843, "dur": 2.305,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997470.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997473.308, "dur": 3.839,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997473.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997477.276, "dur": 29.215,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997477.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507997506.619, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997506.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507997508.763, "dur": 3.456,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997508.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507997512.379, "dur": 7.552,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997512.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507997520.027, "dur": 9.088,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997520.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507997529.244, "dur": 41.056,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997529.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507997570.427, "dur": 8.481,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997570.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997579.035, "dur": 17.089,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997579.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507997596.251, "dur": 15.936,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997596.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997612.315, "dur": 50.913,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997612.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507997663.388, "dur": 13.280,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997663.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997676.795, "dur": 28.289,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997676.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507997705.211, "dur": 21.281,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997705.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507997726.619, "dur": 1.953,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997726.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997728.699, "dur": 4.865,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997728.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997733.691, "dur": 30.816,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997733.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507997764.636, "dur": 1.983,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997764.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507997766.779, "dur": 3.296,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997766.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507997770.204, "dur": 6.496,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997770.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507997776.827, "dur": 8.897,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997776.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507997785.851, "dur": 42.400,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997785.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507997828.380, "dur": 8.479,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997828.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997836.955, "dur": 17.856,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997836.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507997854.940, "dur": 12.735,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997854.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997867.804, "dur": 50.528,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997867.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507997918.459, "dur": 13.344,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997918.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997931.932, "dur": 29.056,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997931.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507997961.115, "dur": 14.337,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997961.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507997975.548, "dur": 1.823,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997975.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997977.500, "dur": 3.840,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997977.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507997981.467, "dur": 29.281,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507997981.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507998010.876, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998010.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507998013.019, "dur": 3.521,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998013.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507998016.668, "dur": 7.488,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998016.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507998024.283, "dur": 9.153,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998024.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507998033.564, "dur": 43.712,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998033.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507998077.403, "dur": 8.737,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998077.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998086.268, "dur": 17.312,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998086.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507998103.707, "dur": 15.936,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998103.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998119.772, "dur": 50.975,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998119.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507998170.876, "dur": 13.088,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998170.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998184.091, "dur": 28.961,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998184.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507998213.180, "dur": 22.560,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998213.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507998235.868, "dur": 2.559,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998235.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998238.556, "dur": 5.728,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998238.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998244.412, "dur": 29.983,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998244.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507998274.524, "dur": 2.080,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998274.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507998276.731, "dur": 3.329,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998276.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507998280.187, "dur": 6.688,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998280.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507998286.972, "dur": 8.864,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998286.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507998295.963, "dur": 42.113,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998295.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507998338.203, "dur": 8.417,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998338.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998346.716, "dur": 17.824,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998346.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507998364.700, "dur": 13.183,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998364.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998378.012, "dur": 50.528,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998378.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507998428.668, "dur": 13.312,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998428.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998442.108, "dur": 29.408,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998442.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507998471.644, "dur": 14.912,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998471.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507998486.683, "dur": 1.825,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998486.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998488.636, "dur": 4.064,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998488.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998492.828, "dur": 29.312,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998492.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507998522.236, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998522.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507998524.412, "dur": 3.391,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998524.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507998527.932, "dur": 7.232,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998527.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507998535.291, "dur": 8.897,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998535.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507998544.316, "dur": 42.432,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998544.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507998586.844, "dur": 8.768,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998586.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998595.740, "dur": 16.512,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998595.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507998612.348, "dur": 16.000,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998612.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998628.508, "dur": 50.400,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998628.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507998679.068, "dur": 13.408,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998679.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998692.604, "dur": 28.544,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998692.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290507998721.276, "dur": 21.984,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998721.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507998743.388, "dur": 2.016,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998743.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998745.500, "dur": 5.280,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998745.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998750.908, "dur": 30.336,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998750.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507998781.372, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998781.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507998783.548, "dur": 3.072,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998783.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507998786.748, "dur": 6.528,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998786.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507998793.404, "dur": 8.960,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998793.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507998802.460, "dur": 41.824,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998802.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507998844.412, "dur": 8.608,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998844.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998853.148, "dur": 17.472,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998853.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290507998870.748, "dur": 13.184,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998870.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998884.060, "dur": 49.568,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998884.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507998933.788, "dur": 13.472,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998933.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998947.388, "dur": 28.224,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998947.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290507998975.740, "dur": 14.208,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998975.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290507998990.076, "dur": 1.824,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998990.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998992.028, "dur": 3.904,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998992.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507998996.028, "dur": 29.632,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507998996.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290507999025.756, "dur": 2.048,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999025.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290507999027.932, "dur": 3.296,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999027.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290507999031.356, "dur": 7.584,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999031.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507999039.036, "dur": 9.056,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999039.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290507999048.220, "dur": 42.080,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999048.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290507999090.396, "dur": 8.480,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999090.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507999099.004, "dur": 16.960,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999099.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290507999116.092, "dur": 16.192,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999116.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507999132.412, "dur": 50.272,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999132.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290507999182.812, "dur": 13.344,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999182.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507999196.284, "dur": 27.808,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999196.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290507999224.188, "dur": 13.376,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999224.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290507999237.660, "dur": 6.240,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999237.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507999244.028, "dur": 147.040,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999244.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507999390.940, "dur": 146.080,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999390.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290507999536.892, "dur": 223.712,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999536.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290507999759.580, "dur": 20.384,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999759.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290507999779.900, "dur": 299.680,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290507999779.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508000081.948, "dur": 0.480,
    "args": {
      "External id": 73,      "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000081.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508000083.452, "dur": 138.784,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000083.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508000225.340, "dur": 1.120,
    "args": {
      "External id": 73,      "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "bytes": 576, "memory bandwidth (GB/s)": 0.5142857142857142
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000225.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508000229.084, "dur": 142.624,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000229.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508000371.612, "dur": 188.672,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000371.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290508000559.260, "dur": 18.336,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000559.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290508000577.500, "dur": 300.064,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000577.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508000879.580, "dur": 0.480,
    "args": {
      "External id": 73,      "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000879.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508000881.052, "dur": 136.800,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508000881.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508001020.956, "dur": 1.312,
    "args": {
      "External id": 73,      "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "bytes": 576, "memory bandwidth (GB/s)": 0.43902439024390244
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508001020.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001023.868, "dur": 139.904,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508001023.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290508001164.092, "dur": 3.360,
    "args": {
      "External id": 73,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504341,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504341, "pid": 0, "tid": 7, "ts": 1290508001164.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507972428.869, "dur": 706.723,
    "args": {
      "External id": 73,      "cbid": 311, "correlation": 504341
    }
  },
  {
    "ph": "s", "id": 504341, "pid": 14565, "tid": 14565, "ts": 1290507972428.869,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290508001169.212, "dur": 1.536,
    "args": {
      "External id": 90,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504354,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504354, "pid": 0, "tid": 7, "ts": 1290508001169.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507974361.446, "dur": 18.342,
    "args": {
      "External id": 90,      "cbid": 211, "correlation": 504354
    }
  },
  {
    "ph": "s", "id": 504354, "pid": 14565, "tid": 14565, "ts": 1290507974361.446,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290508011624.510, "dur": 3.968,
    "args": {
      "External id": 98,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505177,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505177, "pid": 0, "tid": 7, "ts": 1290508011624.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507981596.350, "dur": 13.778,
    "args": {
      "External id": 98,      "cbid": 211, "correlation": 505177
    }
  },
  {
    "ph": "s", "id": 505177, "pid": 14565, "tid": 14565, "ts": 1290507981596.350,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507981617.587, "dur": 0.570,
    "args": {
      "External id": 96,      "cbid": 317, "correlation": 505185
    }
  },
  {
    "ph": "f", "id": 505185, "pid": 14565, "tid": 14565, "ts": 1290507981617.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508011632.126, "dur": 0.832,
    "args": {
      "External id": 96,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505187,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505187, "pid": 0, "tid": 7, "ts": 1290508011632.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508011633.086, "dur": 0.800,
    "args": {
      "External id": 96,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505187,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505187, "pid": 0, "tid": 7, "ts": 1290508011633.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290508011634.046, "dur": 2.080,
    "args": {
      "External id": 96,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505187,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505187, "pid": 0, "tid": 7, "ts": 1290508011634.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290508011636.286, "dur": 6.976,
    "args": {
      "External id": 96,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505187,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505187, "pid": 0, "tid": 7, "ts": 1290508011636.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290508011643.390, "dur": 10.048,
    "args": {
      "External id": 96,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505187,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505187, "pid": 0, "tid": 7, "ts": 1290508011643.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507981618.804, "dur": 19.051,
    "args": {
      "External id": 96,      "cbid": 311, "correlation": 505187
    }
  },
  {
    "ph": "s", "id": 505187, "pid": 14565, "tid": 14565, "ts": 1290507981618.804,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508011654.974, "dur": 1.856,
    "args": {
      "External id": 106,      "device": 0, "context": 1,      "stream": 7, "correlation": 505192,      "bytes": 32, "memory bandwidth (GB/s)": 0.017241379310344827
    }
  },
  {
    "ph": "f", "id": 505192, "pid": 0, "tid": 7, "ts": 1290508011654.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507982164.157, "dur": 18.201,
    "args": {
      "External id": 106,      "cbid": 41, "correlation": 505192
    }
  },
  {
    "ph": "s", "id": 505192, "pid": 14565, "tid": 14565, "ts": 1290507982164.157,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508011657.246, "dur": 1.376,
    "args": {
      "External id": 107,      "device": 0, "context": 1,      "stream": 7, "correlation": 505197,      "bytes": 256, "memory bandwidth (GB/s)": 0.18604651162790697
    }
  },
  {
    "ph": "f", "id": 505197, "pid": 0, "tid": 7, "ts": 1290508011657.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507982187.687, "dur": 4.408,
    "args": {
      "External id": 107,      "cbid": 41, "correlation": 505197
    }
  },
  {
    "ph": "s", "id": 505197, "pid": 14565, "tid": 14565, "ts": 1290507982187.687,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508011659.038, "dur": 1.408,
    "args": {
      "External id": 108,      "device": 0, "context": 1,      "stream": 7, "correlation": 505202,      "bytes": 32, "memory bandwidth (GB/s)": 0.022727272727272728
    }
  },
  {
    "ph": "f", "id": 505202, "pid": 0, "tid": 7, "ts": 1290508011659.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507982194.894, "dur": 3.830,
    "args": {
      "External id": 108,      "cbid": 41, "correlation": 505202
    }
  },
  {
    "ph": "s", "id": 505202, "pid": 14565, "tid": 14565, "ts": 1290507982194.894,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508011660.862, "dur": 1.312,
    "args": {
      "External id": 109,      "device": 0, "context": 1,      "stream": 7, "correlation": 505207,      "bytes": 256, "memory bandwidth (GB/s)": 0.1951219512195122
    }
  },
  {
    "ph": "f", "id": 505207, "pid": 0, "tid": 7, "ts": 1290508011660.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507982201.489, "dur": 3.893,
    "args": {
      "External id": 109,      "cbid": 41, "correlation": 505207
    }
  },
  {
    "ph": "s", "id": 505207, "pid": 14565, "tid": 14565, "ts": 1290507982201.489,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508011662.590, "dur": 1.952,
    "args": {
      "External id": 110,      "device": 0, "context": 1,      "stream": 7, "correlation": 505212,      "bytes": 65536, "memory bandwidth (GB/s)": 33.57377049180328
    }
  },
  {
    "ph": "f", "id": 505212, "pid": 0, "tid": 7, "ts": 1290508011662.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507982208.218, "dur": 3.598,
    "args": {
      "External id": 110,      "cbid": 41, "correlation": 505212
    }
  },
  {
    "ph": "s", "id": 505212, "pid": 14565, "tid": 14565, "ts": 1290507982208.218,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507982219.048, "dur": 0.564,
    "args": {
      "External id": 103,      "cbid": 317, "correlation": 505222
    }
  },
  {
    "ph": "f", "id": 505222, "pid": 14565, "tid": 14565, "ts": 1290507982219.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011670.046, "dur": 1.504,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011670.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011671.678, "dur": 1.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011671.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011673.086, "dur": 1.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011673.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011674.526, "dur": 1.216,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011674.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011676.030, "dur": 1.408,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011676.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011677.598, "dur": 1.152,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011677.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011678.974, "dur": 1.472,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011678.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011680.542, "dur": 1.184,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011680.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011681.854, "dur": 1.440,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011681.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011683.422, "dur": 1.152,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011683.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011684.702, "dur": 1.504,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011684.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011686.334, "dur": 1.184,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011686.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011687.646, "dur": 1.344,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011687.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011689.118, "dur": 1.280,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011689.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011690.526, "dur": 1.248,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011690.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508011691.902, "dur": 1.280,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011691.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508011693.342, "dur": 2.240,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011693.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508011695.710, "dur": 1.856,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011695.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508011697.694, "dur": 1.824,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011697.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508011699.646, "dur": 1.856,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011699.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011701.630, "dur": 2.144,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011701.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011703.998, "dur": 2.176,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011703.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011706.302, "dur": 2.048,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011706.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011708.478, "dur": 2.112,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011708.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011710.718, "dur": 1.824,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011710.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011712.638, "dur": 2.080,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011712.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011714.846, "dur": 1.952,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011714.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011716.926, "dur": 2.016,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011716.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011719.038, "dur": 1.952,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011719.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011721.118, "dur": 2.144,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011721.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011723.390, "dur": 1.600,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011723.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011725.310, "dur": 1.984,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011725.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011727.422, "dur": 1.696,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011727.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011729.246, "dur": 1.760,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011729.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011731.134, "dur": 2.304,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011731.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508011733.566, "dur": 2.016,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011733.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011735.710, "dur": 3.168,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011735.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011739.006, "dur": 3.456,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011739.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011742.590, "dur": 2.304,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011742.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011744.990, "dur": 2.784,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011744.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011747.902, "dur": 2.624,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011747.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011750.622, "dur": 2.144,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011750.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011752.894, "dur": 2.496,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011752.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011755.518, "dur": 2.880,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011755.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011758.494, "dur": 3.232,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011758.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011761.854, "dur": 2.432,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011761.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011764.382, "dur": 2.720,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011764.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011767.230, "dur": 2.880,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011767.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011770.206, "dur": 2.720,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011770.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011773.054, "dur": 2.464,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011773.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011775.678, "dur": 2.784,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011775.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011778.590, "dur": 3.040,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011778.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011781.726, "dur": 2.272,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011781.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011784.094, "dur": 2.560,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011784.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011786.782, "dur": 2.784,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011786.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011789.694, "dur": 2.720,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011789.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011792.542, "dur": 2.304,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011792.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011794.974, "dur": 2.848,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011794.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011797.918, "dur": 2.528,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011797.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011800.542, "dur": 2.752,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011800.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011803.390, "dur": 2.528,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011803.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011806.014, "dur": 3.104,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011806.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011809.214, "dur": 2.304,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011809.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011811.646, "dur": 3.072,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011811.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011814.846, "dur": 3.200,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011814.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011818.174, "dur": 2.688,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011818.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011820.990, "dur": 2.624,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011820.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508011823.742, "dur": 2.656,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011823.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011826.526, "dur": 4.160,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011826.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011830.814, "dur": 4.480,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011830.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011835.390, "dur": 3.936,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011835.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011839.422, "dur": 4.640,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011839.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011844.190, "dur": 4.288,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011844.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011848.574, "dur": 4.128,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011848.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011852.830, "dur": 4.448,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011852.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011857.374, "dur": 4.192,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011857.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011861.662, "dur": 4.448,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011861.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011866.206, "dur": 3.904,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011866.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011870.238, "dur": 4.864,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011870.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011875.230, "dur": 4.064,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011875.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011879.390, "dur": 4.736,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011879.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011884.254, "dur": 4.384,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011884.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011888.734, "dur": 4.384,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011888.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508011893.246, "dur": 4.352,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011893.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290508011897.726, "dur": 7.552,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011897.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290508011905.406, "dur": 7.104,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011905.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011912.606, "dur": 12.416,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011912.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011925.150, "dur": 12.576,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011925.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011937.854, "dur": 12.320,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011937.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011950.270, "dur": 12.608,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011950.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290508011963.038, "dur": 15.968,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011963.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290508011979.134, "dur": 1.280,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011979.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011980.510, "dur": 2.976,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011980.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508011983.614, "dur": 21.856,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508011983.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012005.598, "dur": 29.856,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012005.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508012035.582, "dur": 2.080,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012035.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508012037.790, "dur": 3.488,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012037.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508012041.406, "dur": 7.936,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012041.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508012049.470, "dur": 9.472,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012049.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508012059.038, "dur": 43.008,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012059.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508012102.174, "dur": 8.864,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012102.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012111.166, "dur": 17.504,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012111.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290508012128.798, "dur": 8.544,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012128.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012137.471, "dur": 49.183,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012137.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508012186.782, "dur": 13.632,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012186.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012200.542, "dur": 28.896,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012200.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508012229.566, "dur": 21.537,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012229.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508012251.230, "dur": 9.761,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012251.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508012259.934, "dur": 28.833,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012259.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508012288.894, "dur": 1.984,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012288.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012290.974, "dur": 3.488,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012290.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012294.559, "dur": 29.760,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012294.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508012324.446, "dur": 2.080,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012324.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508012326.622, "dur": 3.296,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012326.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508012330.047, "dur": 6.783,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012330.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508012336.959, "dur": 9.088,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012336.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508012346.143, "dur": 44.671,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012346.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508012390.910, "dur": 8.960,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012390.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012399.998, "dur": 18.592,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012399.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508012418.750, "dur": 16.928,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012418.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012435.838, "dur": 52.161,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012435.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508012488.158, "dur": 13.248,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012488.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012501.566, "dur": 29.536,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012501.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508012531.231, "dur": 14.111,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012531.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508012545.471, "dur": 1.951,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012545.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012547.551, "dur": 3.903,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012547.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012551.550, "dur": 29.856,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012551.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508012581.502, "dur": 2.080,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012581.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508012583.742, "dur": 3.425,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012583.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508012587.327, "dur": 6.623,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012587.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508012594.079, "dur": 9.088,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012594.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508012603.294, "dur": 43.872,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012603.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508012647.294, "dur": 8.960,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012647.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012656.382, "dur": 17.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012656.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508012673.823, "dur": 15.840,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012673.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012689.790, "dur": 50.497,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012689.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508012740.414, "dur": 13.376,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012740.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012753.887, "dur": 29.343,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012753.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508012783.359, "dur": 24.064,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012783.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508012806.431, "dur": 2.047,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012806.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012808.574, "dur": 5.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012808.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012813.982, "dur": 29.729,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012813.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508012843.838, "dur": 2.016,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012843.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508012846.014, "dur": 3.520,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012846.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508012849.663, "dur": 6.719,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012849.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508012856.511, "dur": 9.055,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012856.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508012865.695, "dur": 44.287,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012865.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508012910.142, "dur": 8.608,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012910.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012918.910, "dur": 16.320,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012918.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508012935.359, "dur": 13.184,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012935.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508012948.639, "dur": 50.111,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012948.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508012998.879, "dur": 13.504,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508012998.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013012.510, "dur": 28.033,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013012.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508013040.670, "dur": 14.465,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013040.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508013055.263, "dur": 1.727,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013055.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013057.119, "dur": 3.872,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013057.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013061.087, "dur": 29.824,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013061.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508013091.038, "dur": 2.049,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013091.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508013093.214, "dur": 3.521,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013093.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508013096.863, "dur": 6.624,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013096.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508013103.615, "dur": 9.184,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013103.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508013112.895, "dur": 40.960,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013112.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508013154.015, "dur": 8.608,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013154.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013162.750, "dur": 17.185,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013162.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508013180.031, "dur": 16.160,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013180.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013196.319, "dur": 50.336,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013196.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508013246.783, "dur": 13.280,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013246.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013260.190, "dur": 29.281,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013260.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508013289.599, "dur": 24.351,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013289.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508013312.959, "dur": 2.976,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013312.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013316.063, "dur": 6.944,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013316.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013323.135, "dur": 29.536,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013323.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508013352.799, "dur": 2.048,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013352.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508013354.943, "dur": 3.423,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013354.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508013358.495, "dur": 6.528,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013358.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508013365.119, "dur": 9.120,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013365.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508013374.367, "dur": 42.208,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013374.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508013416.703, "dur": 8.383,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013416.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013425.215, "dur": 17.440,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013425.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508013442.783, "dur": 12.864,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013442.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013455.775, "dur": 50.528,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013455.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508013506.463, "dur": 13.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013506.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013519.903, "dur": 28.960,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013519.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508013548.991, "dur": 14.336,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013548.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508013563.455, "dur": 2.912,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013563.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013566.495, "dur": 3.392,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013566.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013570.015, "dur": 29.856,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013570.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508013599.999, "dur": 2.047,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013599.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508013602.175, "dur": 3.680,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013602.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508013605.951, "dur": 6.752,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013605.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508013612.830, "dur": 9.185,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013612.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508013622.143, "dur": 41.344,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013622.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508013663.615, "dur": 8.640,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013663.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013672.383, "dur": 16.896,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013672.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508013689.439, "dur": 15.904,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013689.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013705.471, "dur": 50.912,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013705.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508013756.543, "dur": 13.216,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013756.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013769.919, "dur": 27.968,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013769.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508013798.015, "dur": 22.912,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013798.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508013821.023, "dur": 2.144,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013821.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013823.295, "dur": 4.832,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013823.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013828.255, "dur": 30.976,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013828.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508013859.359, "dur": 2.048,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013859.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508013861.503, "dur": 3.392,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013861.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508013865.023, "dur": 6.720,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013865.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508013871.839, "dur": 8.960,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013871.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508013880.895, "dur": 41.472,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013880.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508013922.495, "dur": 8.544,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013922.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013931.167, "dur": 17.760,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013931.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508013949.055, "dur": 12.832,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013949.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508013962.015, "dur": 50.848,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508013962.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508014012.991, "dur": 13.536,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014012.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014026.655, "dur": 28.032,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014026.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508014054.783, "dur": 22.208,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014054.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508014077.119, "dur": 8.608,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014077.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290508014085.855, "dur": 27.808,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014085.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508014113.791, "dur": 1.888,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014113.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014115.807, "dur": 3.264,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014115.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014119.199, "dur": 29.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014119.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508014148.639, "dur": 1.984,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014148.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508014150.751, "dur": 3.424,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014150.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508014154.303, "dur": 7.136,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014154.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508014161.567, "dur": 9.184,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014161.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508014170.815, "dur": 41.760,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014170.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508014212.671, "dur": 8.704,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014212.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014221.503, "dur": 17.824,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014221.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508014239.455, "dur": 14.240,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014239.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014253.823, "dur": 51.520,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014253.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508014305.439, "dur": 13.536,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014305.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014319.103, "dur": 28.544,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014319.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508014347.775, "dur": 14.400,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014347.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508014362.303, "dur": 1.632,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014362.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014364.063, "dur": 4.032,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014364.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014368.223, "dur": 29.760,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014368.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508014398.111, "dur": 2.048,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014398.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508014400.287, "dur": 3.456,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014400.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508014403.871, "dur": 7.072,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014403.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508014411.071, "dur": 9.344,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014411.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508014420.511, "dur": 41.856,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014420.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508014462.495, "dur": 8.608,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014462.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014471.263, "dur": 17.152,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014471.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508014488.575, "dur": 15.840,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014488.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014504.543, "dur": 50.880,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014504.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508014555.551, "dur": 13.248,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014555.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014568.927, "dur": 28.288,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014568.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508014597.311, "dur": 21.376,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014597.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508014618.847, "dur": 2.080,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014618.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014621.055, "dur": 4.896,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014621.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014626.047, "dur": 30.752,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014626.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508014656.927, "dur": 2.016,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014656.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508014659.071, "dur": 3.520,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014659.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508014662.751, "dur": 6.496,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014662.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508014669.375, "dur": 8.992,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014669.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508014678.495, "dur": 43.424,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014678.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508014722.047, "dur": 8.352,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014722.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014730.527, "dur": 17.472,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014730.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508014748.127, "dur": 12.800,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014748.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014761.055, "dur": 50.688,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014761.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508014811.871, "dur": 13.248,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014811.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014825.247, "dur": 28.928,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014825.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508014854.303, "dur": 14.432,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014854.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508014868.863, "dur": 1.824,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014868.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014870.815, "dur": 3.776,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014870.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014874.687, "dur": 29.728,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014874.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508014904.543, "dur": 2.016,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014904.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508014906.687, "dur": 3.456,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014906.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508014910.271, "dur": 7.392,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014910.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508014917.791, "dur": 9.120,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014917.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508014927.039, "dur": 43.072,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014927.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508014970.239, "dur": 8.736,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014970.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508014979.103, "dur": 16.928,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014979.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508014996.191, "dur": 15.936,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508014996.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015012.255, "dur": 50.752,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015012.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508015063.135, "dur": 13.248,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015063.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015076.511, "dur": 28.576,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015076.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508015105.215, "dur": 22.112,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015105.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508015127.455, "dur": 3.168,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015127.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015130.751, "dur": 4.864,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015130.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015135.711, "dur": 29.952,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015135.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508015165.791, "dur": 2.080,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015165.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508015167.967, "dur": 3.392,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015167.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508015171.487, "dur": 7.232,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015171.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508015178.847, "dur": 8.736,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015178.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508015187.711, "dur": 42.528,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015187.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508015230.367, "dur": 8.512,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015230.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015238.975, "dur": 17.472,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015238.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508015256.607, "dur": 13.152,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015256.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015269.887, "dur": 50.240,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015269.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508015320.255, "dur": 13.344,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015320.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015333.727, "dur": 29.568,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015333.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508015363.455, "dur": 14.528,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015363.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508015378.111, "dur": 1.760,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015378.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015379.999, "dur": 4.288,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015379.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015384.447, "dur": 29.440,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015384.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508015413.983, "dur": 2.048,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015413.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508015416.127, "dur": 3.424,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015416.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508015419.679, "dur": 6.720,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015419.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508015426.495, "dur": 9.152,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015426.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508015435.775, "dur": 43.424,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015435.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508015479.327, "dur": 8.960,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015479.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015488.383, "dur": 16.320,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015488.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508015504.831, "dur": 16.000,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015504.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015520.991, "dur": 50.944,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015520.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508015572.063, "dur": 13.344,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015572.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015585.535, "dur": 28.800,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015585.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508015614.463, "dur": 22.784,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015614.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508015637.375, "dur": 1.952,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015637.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015639.455, "dur": 5.120,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015639.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015644.703, "dur": 30.240,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015644.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508015675.103, "dur": 2.048,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015675.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508015677.279, "dur": 3.456,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015677.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508015680.863, "dur": 7.520,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015680.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508015688.479, "dur": 9.024,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015688.479,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508015697.599, "dur": 42.304,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015697.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508015740.031, "dur": 8.704,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015740.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015748.863, "dur": 17.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015748.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508015766.303, "dur": 12.800,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015766.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015779.231, "dur": 49.472,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015779.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508015828.863, "dur": 13.312,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015828.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015842.335, "dur": 28.256,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015842.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508015870.719, "dur": 14.272,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015870.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508015885.119, "dur": 1.728,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015885.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015886.975, "dur": 3.840,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015886.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015890.911, "dur": 29.888,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015890.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508015920.927, "dur": 2.048,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015920.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508015923.103, "dur": 3.456,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015923.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508015926.687, "dur": 7.712,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015926.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508015934.495, "dur": 9.056,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015934.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508015943.679, "dur": 41.440,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015943.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508015985.247, "dur": 8.448,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015985.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508015993.855, "dur": 17.248,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508015993.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508016011.231, "dur": 16.128,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016011.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508016027.487, "dur": 50.912,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016027.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508016078.527, "dur": 13.216,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016078.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508016091.871, "dur": 27.872,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016091.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290508016119.871, "dur": 13.248,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016119.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290508016133.247, "dur": 6.336,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016133.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508016139.679, "dur": 148.064,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016139.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508016287.647, "dur": 147.040,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016287.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508016434.591, "dur": 223.136,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016434.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290508016656.543, "dur": 20.544,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016656.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290508016677.023, "dur": 300.000,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016677.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508016979.903, "dur": 0.480,
    "args": {
      "External id": 103,      "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016979.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508016981.407, "dur": 139.264,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508016981.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508017123.391, "dur": 1.217,
    "args": {
      "External id": 103,      "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "bytes": 576, "memory bandwidth (GB/s)": 0.4732949876746097
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017123.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508017127.231, "dur": 142.752,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017127.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508017269.919, "dur": 189.600,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017269.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290508017459.199, "dur": 18.464,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017459.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290508017477.568, "dur": 302.016,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017477.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508017781.727, "dur": 0.513,
    "args": {
      "External id": 103,      "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "bytes": 192, "memory bandwidth (GB/s)": 0.3742690058479532
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017781.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508017783.263, "dur": 138.049,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017783.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508017923.103, "dur": 1.185,
    "args": {
      "External id": 103,      "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "bytes": 576, "memory bandwidth (GB/s)": 0.4860759493670886
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017923.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508017926.015, "dur": 139.809,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508017926.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290508018066.144, "dur": 3.296,
    "args": {
      "External id": 103,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505224,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505224, "pid": 0, "tid": 7, "ts": 1290508018066.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507982220.248, "dur": 383.136,
    "args": {
      "External id": 103,      "cbid": 311, "correlation": 505224
    }
  },
  {
    "ph": "s", "id": 505224, "pid": 14565, "tid": 14565, "ts": 1290507982220.248,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290508018071.360, "dur": 1.568,
    "args": {
      "External id": 120,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505237,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505237, "pid": 0, "tid": 7, "ts": 1290508018071.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507983509.702, "dur": 18.312,
    "args": {
      "External id": 120,      "cbid": 211, "correlation": 505237
    }
  },
  {
    "ph": "s", "id": 505237, "pid": 14565, "tid": 14565, "ts": 1290507983509.702,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290508028528.642, "dur": 3.904,
    "args": {
      "External id": 128,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506060,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506060, "pid": 0, "tid": 7, "ts": 1290508028528.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507990798.104, "dur": 12.840,
    "args": {
      "External id": 128,      "cbid": 211, "correlation": 506060
    }
  },
  {
    "ph": "s", "id": 506060, "pid": 14565, "tid": 14565, "ts": 1290507990798.104,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507990817.640, "dur": 0.573,
    "args": {
      "External id": 126,      "cbid": 317, "correlation": 506068
    }
  },
  {
    "ph": "f", "id": 506068, "pid": 14565, "tid": 14565, "ts": 1290507990817.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508028536.322, "dur": 0.800,
    "args": {
      "External id": 126,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506070,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506070, "pid": 0, "tid": 7, "ts": 1290508028536.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508028537.282, "dur": 0.832,
    "args": {
      "External id": 126,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506070,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506070, "pid": 0, "tid": 7, "ts": 1290508028537.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290508028538.274, "dur": 2.112,
    "args": {
      "External id": 126,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506070,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506070, "pid": 0, "tid": 7, "ts": 1290508028538.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290508028540.514, "dur": 6.784,
    "args": {
      "External id": 126,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506070,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506070, "pid": 0, "tid": 7, "ts": 1290508028540.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290508028547.426, "dur": 11.296,
    "args": {
      "External id": 126,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506070,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506070, "pid": 0, "tid": 7, "ts": 1290508028547.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507990818.857, "dur": 14.798,
    "args": {
      "External id": 126,      "cbid": 311, "correlation": 506070
    }
  },
  {
    "ph": "s", "id": 506070, "pid": 14565, "tid": 14565, "ts": 1290507990818.857,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508028561.154, "dur": 1.376,
    "args": {
      "External id": 136,      "device": 0, "context": 1,      "stream": 7, "correlation": 506075,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 506075, "pid": 0, "tid": 7, "ts": 1290508028561.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507991299.178, "dur": 17.554,
    "args": {
      "External id": 136,      "cbid": 41, "correlation": 506075
    }
  },
  {
    "ph": "s", "id": 506075, "pid": 14565, "tid": 14565, "ts": 1290507991299.178,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508028562.946, "dur": 1.376,
    "args": {
      "External id": 137,      "device": 0, "context": 1,      "stream": 7, "correlation": 506080,      "bytes": 256, "memory bandwidth (GB/s)": 0.18604651162790697
    }
  },
  {
    "ph": "f", "id": 506080, "pid": 0, "tid": 7, "ts": 1290508028562.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507991322.048, "dur": 5.002,
    "args": {
      "External id": 137,      "cbid": 41, "correlation": 506080
    }
  },
  {
    "ph": "s", "id": 506080, "pid": 14565, "tid": 14565, "ts": 1290507991322.048,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508028564.738, "dur": 1.344,
    "args": {
      "External id": 138,      "device": 0, "context": 1,      "stream": 7, "correlation": 506085,      "bytes": 32, "memory bandwidth (GB/s)": 0.023809523809523808
    }
  },
  {
    "ph": "f", "id": 506085, "pid": 0, "tid": 7, "ts": 1290508028564.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507991330.192, "dur": 3.965,
    "args": {
      "External id": 138,      "cbid": 41, "correlation": 506085
    }
  },
  {
    "ph": "s", "id": 506085, "pid": 14565, "tid": 14565, "ts": 1290507991330.192,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508028566.498, "dur": 1.344,
    "args": {
      "External id": 139,      "device": 0, "context": 1,      "stream": 7, "correlation": 506090,      "bytes": 256, "memory bandwidth (GB/s)": 0.19047619047619047
    }
  },
  {
    "ph": "f", "id": 506090, "pid": 0, "tid": 7, "ts": 1290508028566.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507991337.261, "dur": 4.100,
    "args": {
      "External id": 139,      "cbid": 41, "correlation": 506090
    }
  },
  {
    "ph": "s", "id": 506090, "pid": 14565, "tid": 14565, "ts": 1290507991337.261,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508028568.290, "dur": 2.015,
    "args": {
      "External id": 140,      "device": 0, "context": 1,      "stream": 7, "correlation": 506095,      "bytes": 65536, "memory bandwidth (GB/s)": 32.52406947890819
    }
  },
  {
    "ph": "f", "id": 506095, "pid": 0, "tid": 7, "ts": 1290508028568.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507991344.341, "dur": 3.770,
    "args": {
      "External id": 140,      "cbid": 41, "correlation": 506095
    }
  },
  {
    "ph": "s", "id": 506095, "pid": 14565, "tid": 14565, "ts": 1290507991344.341,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507991354.792, "dur": 0.516,
    "args": {
      "External id": 133,      "cbid": 317, "correlation": 506105
    }
  },
  {
    "ph": "f", "id": 506105, "pid": 14565, "tid": 14565, "ts": 1290507991354.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028575.202, "dur": 1.504,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028575.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028576.834, "dur": 1.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028576.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028578.242, "dur": 1.312,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028578.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028579.682, "dur": 1.152,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028579.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028581.058, "dur": 1.440,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028581.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028582.626, "dur": 1.184,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028582.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028584.034, "dur": 1.440,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028584.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028585.570, "dur": 1.184,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028585.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028586.850, "dur": 1.472,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028586.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028588.450, "dur": 1.152,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028588.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028589.730, "dur": 1.376,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028589.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028591.202, "dur": 1.184,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028591.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028592.514, "dur": 1.248,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028592.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028593.858, "dur": 1.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028593.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028595.266, "dur": 1.248,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028595.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508028596.642, "dur": 1.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028596.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508028598.050, "dur": 1.824,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028598.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508028600.001, "dur": 1.889,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028600.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508028602.050, "dur": 1.952,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028602.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508028604.130, "dur": 1.792,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028604.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028606.082, "dur": 2.176,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028606.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028608.386, "dur": 2.336,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028608.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028610.850, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028610.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028612.994, "dur": 1.952,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028612.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028615.138, "dur": 2.080,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028615.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028617.346, "dur": 2.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028617.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028620.034, "dur": 1.984,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028620.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028622.146, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028622.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028624.418, "dur": 1.920,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028624.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028626.434, "dur": 1.760,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028626.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028628.322, "dur": 1.600,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028628.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028630.242, "dur": 1.888,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028630.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028632.258, "dur": 1.568,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028632.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028633.986, "dur": 1.600,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028633.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028635.714, "dur": 1.952,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028635.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508028637.794, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028637.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028639.906, "dur": 3.232,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028639.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028643.266, "dur": 3.008,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028643.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028646.370, "dur": 2.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028646.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028649.090, "dur": 2.688,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028649.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028651.906, "dur": 2.656,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028651.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028654.690, "dur": 2.464,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028654.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028657.250, "dur": 2.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028657.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028659.970, "dur": 2.784,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028659.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028662.882, "dur": 2.784,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028662.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028665.762, "dur": 2.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028665.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028668.482, "dur": 2.848,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028668.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028671.458, "dur": 2.816,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028671.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028674.402, "dur": 2.944,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028674.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028677.442, "dur": 2.496,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028677.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028680.066, "dur": 2.848,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028680.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028683.042, "dur": 3.232,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028683.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028686.402, "dur": 2.208,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028686.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028688.738, "dur": 2.304,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028688.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028691.170, "dur": 2.560,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028691.170,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028693.858, "dur": 2.720,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028693.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028696.674, "dur": 2.272,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028696.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028699.074, "dur": 2.784,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028699.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028701.986, "dur": 2.976,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028701.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028705.058, "dur": 2.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028705.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028707.778, "dur": 2.464,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028707.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028710.338, "dur": 2.624,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028710.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028713.090, "dur": 2.432,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028713.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028715.650, "dur": 3.040,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028715.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028718.818, "dur": 2.816,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028718.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028721.762, "dur": 2.848,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028721.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028724.770, "dur": 2.624,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028724.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508028727.554, "dur": 2.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028727.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028730.242, "dur": 4.352,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028730.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028734.722, "dur": 4.576,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028734.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028739.394, "dur": 4.128,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028739.394,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028743.618, "dur": 4.544,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028743.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028748.258, "dur": 4.384,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028748.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028752.738, "dur": 4.160,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028752.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028757.026, "dur": 4.448,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028757.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028761.602, "dur": 4.256,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028761.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028765.954, "dur": 4.096,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028765.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028770.146, "dur": 3.840,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028770.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028774.082, "dur": 4.416,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028774.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028778.626, "dur": 4.064,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028778.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028782.786, "dur": 4.384,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028782.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028787.266, "dur": 4.960,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028787.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028792.322, "dur": 4.128,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028792.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508028796.546, "dur": 4.704,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028796.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290508028801.378, "dur": 7.584,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028801.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290508028809.058, "dur": 6.816,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028809.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028816.002, "dur": 12.480,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028816.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028828.578, "dur": 12.544,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028828.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028841.250, "dur": 12.192,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028841.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028853.570, "dur": 12.544,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028853.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290508028866.274, "dur": 15.424,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028866.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290508028881.826, "dur": 1.248,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028881.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028883.202, "dur": 2.944,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028883.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508028886.306, "dur": 21.792,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028886.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028908.226, "dur": 29.984,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028908.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508028938.338, "dur": 2.080,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028938.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508028940.514, "dur": 3.488,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028940.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508028944.130, "dur": 6.624,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028944.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508028950.850, "dur": 9.728,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028950.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508028960.674, "dur": 41.408,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508028960.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508029002.210, "dur": 8.640,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029002.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029010.978, "dur": 17.344,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029010.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290508029028.450, "dur": 9.216,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029028.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029037.794, "dur": 49.216,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029037.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508029087.170, "dur": 13.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029087.170,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029100.578, "dur": 28.864,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029100.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508029129.602, "dur": 21.728,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029129.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508029151.490, "dur": 9.472,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029151.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508029159.938, "dur": 27.616,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029159.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508029187.682, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029187.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029189.794, "dur": 4.032,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029189.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029193.922, "dur": 29.728,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029193.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508029223.746, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029223.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508029225.890, "dur": 3.392,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029225.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508029229.410, "dur": 6.912,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029229.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508029236.450, "dur": 8.960,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029236.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508029245.506, "dur": 43.936,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029245.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508029289.570, "dur": 9.120,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029289.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029298.850, "dur": 17.376,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029298.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508029316.386, "dur": 17.792,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029316.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029334.306, "dur": 51.456,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029334.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508029385.922, "dur": 13.312,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029385.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029399.362, "dur": 29.248,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029399.362,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508029428.738, "dur": 14.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029428.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508029442.914, "dur": 1.792,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029442.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029444.834, "dur": 3.840,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029444.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029448.802, "dur": 29.664,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029448.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508029478.562, "dur": 2.080,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029478.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508029480.738, "dur": 3.424,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029480.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508029484.290, "dur": 6.816,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029484.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508029491.234, "dur": 9.152,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029491.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508029500.514, "dur": 42.656,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029500.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508029543.330, "dur": 8.864,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029543.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029552.322, "dur": 16.992,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029552.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508029569.442, "dur": 15.712,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029569.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029585.282, "dur": 51.520,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029585.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508029636.930, "dur": 13.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029636.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029650.306, "dur": 28.832,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029650.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508029679.266, "dur": 23.264,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029679.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508029702.658, "dur": 2.112,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029702.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029704.866, "dur": 5.184,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029704.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029710.114, "dur": 29.696,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029710.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508029739.938, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029739.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508029742.082, "dur": 3.712,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029742.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508029745.922, "dur": 6.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029745.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508029752.610, "dur": 8.928,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029752.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508029761.666, "dur": 42.304,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029761.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508029804.098, "dur": 8.544,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029804.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029812.802, "dur": 15.968,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029812.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508029828.930, "dur": 12.992,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029828.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029842.050, "dur": 51.232,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029842.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508029893.378, "dur": 13.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029893.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029906.786, "dur": 28.064,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029906.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508029934.978, "dur": 14.368,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029934.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508029949.474, "dur": 1.728,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029949.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029951.330, "dur": 3.776,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029951.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508029955.234, "dur": 29.344,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029955.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508029984.706, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029984.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508029986.914, "dur": 3.456,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029986.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508029990.530, "dur": 6.784,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029990.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508029997.442, "dur": 9.024,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508029997.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508030006.594, "dur": 41.728,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030006.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508030048.450, "dur": 8.448,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030048.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030057.026, "dur": 17.056,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030057.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508030074.210, "dur": 15.968,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030074.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030090.306, "dur": 50.336,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030090.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508030140.770, "dur": 13.120,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030140.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030154.018, "dur": 29.664,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030154.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508030183.810, "dur": 24.096,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030183.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508030206.850, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030206.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030208.994, "dur": 4.448,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030208.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030213.570, "dur": 31.136,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030213.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508030244.834, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030244.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508030247.010, "dur": 3.424,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030247.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508030250.562, "dur": 6.912,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030250.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508030257.570, "dur": 8.960,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030257.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508030266.658, "dur": 42.816,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030266.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508030309.602, "dur": 8.512,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030309.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030318.242, "dur": 17.472,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030318.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508030335.842, "dur": 12.864,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030335.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030348.834, "dur": 50.144,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030348.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508030399.138, "dur": 13.312,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030399.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030412.578, "dur": 28.704,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030412.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508030441.410, "dur": 14.432,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030441.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508030455.970, "dur": 1.728,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030455.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030457.826, "dur": 3.840,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030457.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030461.794, "dur": 29.376,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030461.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508030491.266, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030491.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508030493.442, "dur": 3.488,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030493.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508030497.026, "dur": 6.976,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030497.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508030504.130, "dur": 9.152,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030504.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508030513.378, "dur": 42.944,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030513.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508030556.450, "dur": 8.672,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030556.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030565.282, "dur": 17.216,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030565.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508030582.626, "dur": 15.712,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030582.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030598.466, "dur": 50.880,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030598.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508030649.506, "dur": 13.216,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030649.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030662.882, "dur": 27.488,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030662.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508030690.498, "dur": 22.912,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030690.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508030713.506, "dur": 2.080,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030713.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030715.714, "dur": 4.896,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030715.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030720.738, "dur": 30.688,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030720.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508030751.554, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030751.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508030753.666, "dur": 3.392,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030753.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508030757.186, "dur": 7.584,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030757.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508030764.866, "dur": 8.896,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030764.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508030773.826, "dur": 41.856,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030773.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508030815.810, "dur": 8.544,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030815.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030824.514, "dur": 17.408,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030824.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508030842.050, "dur": 12.928,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030842.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030855.106, "dur": 51.008,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030855.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508030906.274, "dur": 13.376,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030906.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508030919.778, "dur": 27.904,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030919.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508030947.810, "dur": 22.208,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030947.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508030970.146, "dur": 8.384,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030970.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290508030978.658, "dur": 27.296,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508030978.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508031006.082, "dur": 1.888,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031006.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031008.098, "dur": 3.520,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031008.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031011.746, "dur": 29.664,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031011.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508031041.538, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031041.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508031043.714, "dur": 3.424,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031043.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508031047.266, "dur": 7.296,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031047.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508031054.658, "dur": 9.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031054.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508031064.034, "dur": 42.080,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031064.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508031106.242, "dur": 8.768,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031106.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031115.138, "dur": 17.536,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031115.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508031132.802, "dur": 13.952,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031132.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031146.882, "dur": 51.872,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031146.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508031198.882, "dur": 13.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031198.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031212.258, "dur": 28.064,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031212.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508031240.450, "dur": 14.304,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031240.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508031254.914, "dur": 1.792,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031254.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031256.834, "dur": 3.840,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031256.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031260.802, "dur": 29.504,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031260.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508031290.402, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031290.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508031292.546, "dur": 3.456,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031292.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508031296.130, "dur": 6.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031296.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508031302.850, "dur": 9.120,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031302.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508031312.066, "dur": 42.304,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031312.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508031354.498, "dur": 8.320,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031354.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031362.978, "dur": 17.056,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031362.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508031380.194, "dur": 15.968,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031380.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031396.290, "dur": 50.400,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031396.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508031446.818, "dur": 13.184,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031446.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031460.130, "dur": 28.736,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031460.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508031488.994, "dur": 21.376,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031488.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508031510.498, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031510.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031512.674, "dur": 4.832,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031512.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031517.634, "dur": 31.136,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031517.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508031548.898, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031548.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508031551.042, "dur": 3.360,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031551.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508031554.562, "dur": 6.560,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031554.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508031561.250, "dur": 8.992,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031561.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508031570.370, "dur": 40.672,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031570.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508031611.138, "dur": 8.192,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031611.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031619.458, "dur": 17.728,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031619.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508031637.314, "dur": 13.024,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031637.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031650.434, "dur": 50.624,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031650.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508031701.186, "dur": 13.376,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031701.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031714.690, "dur": 28.768,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031714.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508031743.554, "dur": 14.432,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031743.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508031758.082, "dur": 1.856,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031758.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031760.066, "dur": 3.744,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031760.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031763.938, "dur": 29.440,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031763.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508031793.506, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031793.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508031795.650, "dur": 3.360,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031795.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508031799.138, "dur": 7.616,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031799.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508031806.850, "dur": 9.120,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031806.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508031816.098, "dur": 41.312,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031816.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508031857.538, "dur": 8.576,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031857.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031866.242, "dur": 17.088,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031866.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508031883.458, "dur": 15.904,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031883.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031899.458, "dur": 50.976,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031899.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508031950.562, "dur": 13.248,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031950.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508031963.938, "dur": 28.672,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031963.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508031992.738, "dur": 22.080,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508031992.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508032014.978, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032014.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032017.154, "dur": 6.784,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032017.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032024.066, "dur": 30.017,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032024.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508032054.210, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032054.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508032056.354, "dur": 2.944,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032056.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508032059.426, "dur": 6.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032059.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508032066.147, "dur": 8.927,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032066.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508032075.202, "dur": 40.960,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032075.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508032116.290, "dur": 8.416,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032116.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032124.834, "dur": 17.472,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032124.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508032142.466, "dur": 13.120,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032142.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032155.746, "dur": 50.144,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032155.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508032206.019, "dur": 13.375,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032206.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032219.490, "dur": 29.472,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032219.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508032249.122, "dur": 14.560,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032249.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508032263.810, "dur": 2.721,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032263.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032266.691, "dur": 3.359,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032266.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032270.178, "dur": 29.408,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032270.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508032299.714, "dur": 2.048,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032299.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508032301.859, "dur": 3.007,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032301.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508032304.962, "dur": 6.625,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032304.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508032311.683, "dur": 9.088,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032311.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508032320.866, "dur": 42.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032320.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508032363.011, "dur": 8.639,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032363.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032371.746, "dur": 16.576,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032371.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508032388.418, "dur": 15.904,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032388.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032404.482, "dur": 50.304,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032404.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508032454.946, "dur": 13.280,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032454.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032468.355, "dur": 28.576,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032468.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508032497.058, "dur": 21.696,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032497.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508032518.883, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032518.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032521.026, "dur": 4.801,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032521.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032525.954, "dur": 30.880,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032525.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508032556.963, "dur": 2.047,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032556.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508032559.139, "dur": 3.327,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032559.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508032562.595, "dur": 7.328,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032562.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508032570.050, "dur": 9.088,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032570.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508032579.235, "dur": 42.463,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032579.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508032621.858, "dur": 8.577,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032621.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032630.562, "dur": 17.665,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032630.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508032648.387, "dur": 12.800,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032648.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032661.315, "dur": 49.760,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032661.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508032711.235, "dur": 13.375,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032711.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032724.739, "dur": 28.031,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032724.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508032752.899, "dur": 14.112,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032752.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508032767.139, "dur": 1.823,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032767.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032769.091, "dur": 3.840,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032769.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032773.027, "dur": 29.472,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032773.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508032802.627, "dur": 2.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032802.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508032804.770, "dur": 3.329,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032804.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508032808.226, "dur": 6.592,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032808.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508032814.947, "dur": 8.960,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032814.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508032824.035, "dur": 43.456,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032824.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508032867.618, "dur": 8.481,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032867.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032876.259, "dur": 17.312,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032876.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508032893.698, "dur": 16.033,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032893.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032909.858, "dur": 51.649,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032909.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508032961.635, "dur": 13.312,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032961.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508032975.043, "dur": 27.871,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508032975.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290508033003.043, "dur": 13.088,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033003.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290508033016.258, "dur": 6.368,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033016.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508033022.755, "dur": 147.584,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033022.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508033170.210, "dur": 145.281,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033170.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508033315.427, "dur": 220.928,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033315.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290508033535.331, "dur": 21.856,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033535.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290508033557.122, "dur": 300.769,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033557.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508033859.491, "dur": 0.480,
    "args": {
      "External id": 133,      "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033859.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508033860.963, "dur": 138.496,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508033860.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508034000.963, "dur": 1.152,
    "args": {
      "External id": 133,      "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "bytes": 576, "memory bandwidth (GB/s)": 0.5
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034000.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508034004.707, "dur": 140.800,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034004.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508034145.411, "dur": 185.824,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034145.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290508034330.179, "dur": 20.064,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034330.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290508034350.211, "dur": 302.016,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034350.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508034655.203, "dur": 0.480,
    "args": {
      "External id": 133,      "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034655.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508034656.707, "dur": 137.760,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034656.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508034796.515, "dur": 1.248,
    "args": {
      "External id": 133,      "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "bytes": 576, "memory bandwidth (GB/s)": 0.46153846153846156
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034796.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508034799.459, "dur": 140.864,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034799.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290508034940.643, "dur": 3.360,
    "args": {
      "External id": 133,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506107,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506107, "pid": 0, "tid": 7, "ts": 1290508034940.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507991355.930, "dur": 408.263,
    "args": {
      "External id": 133,      "cbid": 311, "correlation": 506107
    }
  },
  {
    "ph": "s", "id": 506107, "pid": 14565, "tid": 14565, "ts": 1290507991355.930,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290508034946.883, "dur": 1.536,
    "args": {
      "External id": 150,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506120,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506120, "pid": 0, "tid": 7, "ts": 1290508034946.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507992518.452, "dur": 12.728,
    "args": {
      "External id": 150,      "cbid": 211, "correlation": 506120
    }
  },
  {
    "ph": "s", "id": 506120, "pid": 14565, "tid": 14565, "ts": 1290507992518.452,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290508045401.925, "dur": 3.936,
    "args": {
      "External id": 158,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506943,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506943, "pid": 0, "tid": 7, "ts": 1290508045401.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507997993.005, "dur": 14.044,
    "args": {
      "External id": 158,      "cbid": 211, "correlation": 506943
    }
  },
  {
    "ph": "s", "id": 506943, "pid": 14565, "tid": 14565, "ts": 1290507997993.005,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507998014.142, "dur": 0.674,
    "args": {
      "External id": 156,      "cbid": 317, "correlation": 506951
    }
  },
  {
    "ph": "f", "id": 506951, "pid": 14565, "tid": 14565, "ts": 1290507998014.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508045409.861, "dur": 0.832,
    "args": {
      "External id": 156,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506953,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506953, "pid": 0, "tid": 7, "ts": 1290508045409.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508045410.821, "dur": 0.800,
    "args": {
      "External id": 156,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506953,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506953, "pid": 0, "tid": 7, "ts": 1290508045410.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290508045411.781, "dur": 2.016,
    "args": {
      "External id": 156,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506953,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506953, "pid": 0, "tid": 7, "ts": 1290508045411.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290508045413.957, "dur": 6.944,
    "args": {
      "External id": 156,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506953,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506953, "pid": 0, "tid": 7, "ts": 1290508045413.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290508045421.029, "dur": 11.488,
    "args": {
      "External id": 156,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506953,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506953, "pid": 0, "tid": 7, "ts": 1290508045421.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507998015.445, "dur": 13.804,
    "args": {
      "External id": 156,      "cbid": 311, "correlation": 506953
    }
  },
  {
    "ph": "s", "id": 506953, "pid": 14565, "tid": 14565, "ts": 1290507998015.445,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508045434.629, "dur": 1.440,
    "args": {
      "External id": 166,      "device": 0, "context": 1,      "stream": 7, "correlation": 506958,      "bytes": 32, "memory bandwidth (GB/s)": 0.022222222222222223
    }
  },
  {
    "ph": "f", "id": 506958, "pid": 0, "tid": 7, "ts": 1290508045434.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507998458.700, "dur": 18.961,
    "args": {
      "External id": 166,      "cbid": 41, "correlation": 506958
    }
  },
  {
    "ph": "s", "id": 506958, "pid": 14565, "tid": 14565, "ts": 1290507998458.700,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508045436.453, "dur": 1.312,
    "args": {
      "External id": 167,      "device": 0, "context": 1,      "stream": 7, "correlation": 506963,      "bytes": 256, "memory bandwidth (GB/s)": 0.1951219512195122
    }
  },
  {
    "ph": "f", "id": 506963, "pid": 0, "tid": 7, "ts": 1290508045436.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507998482.870, "dur": 5.067,
    "args": {
      "External id": 167,      "cbid": 41, "correlation": 506963
    }
  },
  {
    "ph": "s", "id": 506963, "pid": 14565, "tid": 14565, "ts": 1290507998482.870,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508045438.213, "dur": 1.344,
    "args": {
      "External id": 168,      "device": 0, "context": 1,      "stream": 7, "correlation": 506968,      "bytes": 32, "memory bandwidth (GB/s)": 0.023809523809523808
    }
  },
  {
    "ph": "f", "id": 506968, "pid": 0, "tid": 7, "ts": 1290508045438.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507998491.139, "dur": 4.090,
    "args": {
      "External id": 168,      "cbid": 41, "correlation": 506968
    }
  },
  {
    "ph": "s", "id": 506968, "pid": 14565, "tid": 14565, "ts": 1290507998491.139,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508045440.005, "dur": 1.408,
    "args": {
      "External id": 169,      "device": 0, "context": 1,      "stream": 7, "correlation": 506973,      "bytes": 256, "memory bandwidth (GB/s)": 0.18181818181818182
    }
  },
  {
    "ph": "f", "id": 506973, "pid": 0, "tid": 7, "ts": 1290508045440.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507998498.464, "dur": 4.844,
    "args": {
      "External id": 169,      "cbid": 41, "correlation": 506973
    }
  },
  {
    "ph": "s", "id": 506973, "pid": 14565, "tid": 14565, "ts": 1290507998498.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508045441.829, "dur": 2.048,
    "args": {
      "External id": 170,      "device": 0, "context": 1,      "stream": 7, "correlation": 506978,      "bytes": 65536, "memory bandwidth (GB/s)": 32
    }
  },
  {
    "ph": "f", "id": 506978, "pid": 0, "tid": 7, "ts": 1290508045441.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290507998506.383, "dur": 3.805,
    "args": {
      "External id": 170,      "cbid": 41, "correlation": 506978
    }
  },
  {
    "ph": "s", "id": 506978, "pid": 14565, "tid": 14565, "ts": 1290507998506.383,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290507998517.183, "dur": 0.696,
    "args": {
      "External id": 163,      "cbid": 317, "correlation": 506988
    }
  },
  {
    "ph": "f", "id": 506988, "pid": 14565, "tid": 14565, "ts": 1290507998517.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045450.693, "dur": 1.504,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045450.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045452.325, "dur": 1.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045452.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045453.733, "dur": 1.376,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045453.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045455.237, "dur": 1.376,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045455.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045456.741, "dur": 1.184,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045456.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045458.053, "dur": 1.152,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045458.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045459.333, "dur": 1.440,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045459.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045460.901, "dur": 1.152,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045460.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045462.213, "dur": 1.312,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045462.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045463.653, "dur": 1.152,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045463.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045464.933, "dur": 1.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045464.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045466.309, "dur": 1.216,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045466.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045467.621, "dur": 1.216,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045467.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045468.965, "dur": 1.216,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045468.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045470.309, "dur": 1.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045470.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508045471.845, "dur": 1.440,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045471.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508045473.445, "dur": 1.824,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045473.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508045475.397, "dur": 2.144,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045475.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508045477.669, "dur": 2.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045477.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508045479.845, "dur": 2.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045479.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045481.989, "dur": 2.080,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045481.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045484.293, "dur": 2.208,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045484.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045486.629, "dur": 2.048,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045486.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045488.805, "dur": 1.952,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045488.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045490.853, "dur": 1.760,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045490.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045492.709, "dur": 1.920,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045492.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045494.757, "dur": 2.592,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045494.757,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045497.477, "dur": 1.984,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045497.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045499.749, "dur": 1.920,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045499.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045501.765, "dur": 1.760,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045501.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045503.621, "dur": 1.600,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045503.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045505.541, "dur": 1.920,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045505.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045507.557, "dur": 1.632,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045507.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045509.317, "dur": 1.632,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045509.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045511.077, "dur": 1.984,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045511.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508045513.189, "dur": 2.528,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045513.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045515.845, "dur": 3.328,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045515.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045519.301, "dur": 2.976,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045519.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045522.405, "dur": 2.368,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045522.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045524.901, "dur": 2.816,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045524.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045527.845, "dur": 2.592,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045527.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045530.565, "dur": 2.272,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045530.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045532.933, "dur": 2.816,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045532.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045535.877, "dur": 2.880,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045535.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045538.885, "dur": 2.848,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045538.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045541.861, "dur": 2.656,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045541.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045544.645, "dur": 2.848,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045544.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045547.621, "dur": 2.816,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045547.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045550.565, "dur": 3.232,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045550.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045553.893, "dur": 2.496,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045553.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045556.517, "dur": 2.720,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045556.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045559.397, "dur": 3.200,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045559.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045562.725, "dur": 2.272,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045562.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045565.093, "dur": 2.336,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045565.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045567.557, "dur": 2.496,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045567.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045570.149, "dur": 3.072,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045570.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045573.349, "dur": 2.240,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045573.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045575.717, "dur": 2.848,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045575.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045578.693, "dur": 2.880,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045578.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045581.701, "dur": 2.656,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045581.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045584.453, "dur": 2.432,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045584.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045587.013, "dur": 2.656,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045587.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045589.797, "dur": 2.464,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045589.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045592.357, "dur": 3.040,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045592.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045595.557, "dur": 2.784,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045595.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045598.469, "dur": 3.040,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045598.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045601.669, "dur": 2.624,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045601.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508045604.453, "dur": 2.688,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045604.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045607.269, "dur": 4.448,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045607.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045611.845, "dur": 4.512,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045611.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045616.453, "dur": 4.128,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045616.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045620.677, "dur": 4.800,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045620.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045625.573, "dur": 4.480,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045625.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045630.149, "dur": 4.032,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045630.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045634.309, "dur": 4.384,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045634.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045638.821, "dur": 4.352,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045638.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045643.269, "dur": 4.256,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045643.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045647.621, "dur": 3.840,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045647.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045651.589, "dur": 4.512,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045651.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045656.229, "dur": 3.904,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045656.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045660.229, "dur": 4.288,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045660.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045664.645, "dur": 4.896,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045664.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045669.669, "dur": 4.032,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045669.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508045673.797, "dur": 4.736,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045673.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290508045678.661, "dur": 7.552,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045678.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290508045686.341, "dur": 6.784,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045686.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045693.253, "dur": 12.480,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045693.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045705.861, "dur": 12.640,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045705.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045718.629, "dur": 12.320,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045718.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045731.077, "dur": 12.480,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045731.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290508045743.717, "dur": 15.648,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045743.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290508045759.493, "dur": 1.248,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045759.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045760.869, "dur": 3.008,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045760.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508045764.037, "dur": 21.664,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045764.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045785.829, "dur": 29.568,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045785.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508045815.525, "dur": 2.048,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045815.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508045817.701, "dur": 3.456,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045817.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508045821.285, "dur": 6.752,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045821.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508045828.133, "dur": 9.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045828.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508045837.541, "dur": 41.184,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045837.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508045878.853, "dur": 8.576,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045878.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045887.557, "dur": 17.440,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045887.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290508045905.157, "dur": 8.576,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045905.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045913.861, "dur": 49.152,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045913.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508045963.141, "dur": 13.376,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045963.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045976.677, "dur": 29.504,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508045976.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508046006.341, "dur": 22.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046006.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508046028.485, "dur": 9.792,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046028.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508046037.189, "dur": 28.704,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046037.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508046066.021, "dur": 2.624,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046066.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046068.741, "dur": 3.296,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046068.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046072.133, "dur": 29.952,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046072.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508046102.213, "dur": 2.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046102.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508046104.357, "dur": 3.360,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046104.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508046107.845, "dur": 6.848,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046107.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508046114.789, "dur": 8.928,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046114.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508046123.845, "dur": 41.792,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046123.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508046165.765, "dur": 8.864,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046165.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046174.789, "dur": 17.696,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046174.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508046192.645, "dur": 17.600,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046192.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046210.405, "dur": 50.880,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046210.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508046261.445, "dur": 13.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046261.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046274.853, "dur": 28.896,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046274.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508046303.877, "dur": 14.144,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046303.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508046318.149, "dur": 1.728,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046318.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046320.037, "dur": 3.840,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046320.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046323.973, "dur": 29.856,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046323.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508046353.925, "dur": 2.048,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046353.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508046356.101, "dur": 3.488,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046356.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508046359.717, "dur": 7.392,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046359.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508046367.237, "dur": 9.024,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046367.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508046376.389, "dur": 44.576,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046376.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508046421.093, "dur": 8.960,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046421.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046430.181, "dur": 17.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046430.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508046447.589, "dur": 15.712,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046447.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046463.429, "dur": 51.360,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046463.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508046514.917, "dur": 13.312,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046514.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046528.325, "dur": 29.088,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046528.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508046557.541, "dur": 23.968,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046557.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508046581.637, "dur": 3.648,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046581.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046584.261, "dur": 4.960,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046584.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046589.317, "dur": 29.536,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046589.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508046618.981, "dur": 2.048,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046618.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508046621.157, "dur": 3.456,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046621.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508046624.741, "dur": 6.560,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046624.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508046631.397, "dur": 9.088,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046631.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508046640.613, "dur": 41.216,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046640.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508046681.989, "dur": 8.544,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046681.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046690.661, "dur": 16.192,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046690.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508046706.981, "dur": 13.024,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046706.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046720.133, "dur": 50.176,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046720.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508046770.437, "dur": 13.344,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046770.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046783.909, "dur": 28.000,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046783.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508046812.037, "dur": 14.496,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046812.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508046826.661, "dur": 1.760,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046826.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046828.549, "dur": 4.032,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046828.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046832.709, "dur": 29.472,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046832.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508046862.309, "dur": 2.048,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046862.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508046864.485, "dur": 3.680,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046864.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508046868.293, "dur": 6.752,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046868.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508046875.173, "dur": 9.088,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046875.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508046884.389, "dur": 43.136,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046884.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508046927.653, "dur": 8.737,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046927.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046936.517, "dur": 17.120,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046936.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508046953.765, "dur": 15.808,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046953.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508046969.701, "dur": 50.400,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508046969.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508047020.229, "dur": 13.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047020.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047033.637, "dur": 29.856,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047033.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508047063.622, "dur": 24.735,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047063.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508047087.269, "dur": 1.952,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047087.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047089.349, "dur": 5.216,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047089.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047094.693, "dur": 30.176,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047094.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508047124.965, "dur": 2.048,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047124.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508047127.141, "dur": 3.392,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047127.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508047130.661, "dur": 6.624,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047130.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508047137.381, "dur": 8.961,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047137.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508047146.469, "dur": 42.176,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047146.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508047188.773, "dur": 8.416,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047188.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047197.317, "dur": 17.376,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047197.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508047214.821, "dur": 12.801,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047214.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047227.749, "dur": 50.272,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047227.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508047278.181, "dur": 13.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047278.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047291.589, "dur": 28.640,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047291.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508047320.357, "dur": 14.400,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047320.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508047334.853, "dur": 1.792,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047334.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047336.773, "dur": 4.256,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047336.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047341.125, "dur": 29.568,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047341.125,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508047370.790, "dur": 2.047,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047370.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508047372.933, "dur": 3.424,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047372.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508047376.485, "dur": 6.752,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047376.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508047383.397, "dur": 9.184,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047383.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508047392.678, "dur": 43.071,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047392.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508047435.878, "dur": 8.671,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047435.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047444.677, "dur": 16.704,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047444.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508047461.509, "dur": 15.648,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047461.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047477.285, "dur": 50.977,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047477.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508047528.422, "dur": 13.056,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047528.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047541.605, "dur": 27.617,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047541.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508047569.318, "dur": 22.431,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047569.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508047591.846, "dur": 3.072,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047591.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047595.013, "dur": 4.928,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047595.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047600.037, "dur": 31.073,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047600.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508047631.206, "dur": 2.047,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047631.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508047633.382, "dur": 3.327,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047633.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508047636.805, "dur": 6.656,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047636.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508047643.557, "dur": 9.088,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047643.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508047652.741, "dur": 42.433,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047652.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508047695.301, "dur": 8.481,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047695.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047703.942, "dur": 17.728,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047703.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508047721.798, "dur": 12.896,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047721.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047734.854, "dur": 50.784,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047734.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508047785.765, "dur": 13.408,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047785.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047799.269, "dur": 28.513,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047799.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508047827.909, "dur": 22.464,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047827.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508047850.502, "dur": 8.639,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047850.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290508047859.270, "dur": 28.736,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047859.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508047888.134, "dur": 1.823,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047888.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047890.086, "dur": 3.232,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047890.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047893.446, "dur": 29.536,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047893.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508047923.142, "dur": 2.079,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047923.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508047925.350, "dur": 3.487,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047925.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508047928.966, "dur": 6.592,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047928.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508047935.654, "dur": 9.279,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047935.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508047945.030, "dur": 43.743,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047945.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508047988.902, "dur": 8.832,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047988.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508047997.862, "dur": 17.984,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508047997.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508048015.973, "dur": 15.265,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048015.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048031.365, "dur": 51.168,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048031.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508048082.630, "dur": 13.376,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048082.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048096.133, "dur": 28.513,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048096.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508048124.806, "dur": 14.752,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048124.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508048139.686, "dur": 1.728,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048139.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048141.542, "dur": 4.032,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048141.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048145.701, "dur": 29.536,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048145.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508048175.366, "dur": 2.047,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048175.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508048177.542, "dur": 3.520,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048177.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508048181.189, "dur": 6.688,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048181.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508048188.006, "dur": 9.376,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048188.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508048197.510, "dur": 41.791,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048197.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508048239.398, "dur": 8.640,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048239.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048248.198, "dur": 17.087,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048248.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508048265.414, "dur": 15.904,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048265.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048281.446, "dur": 51.519,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048281.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508048333.094, "dur": 13.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048333.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048346.502, "dur": 28.448,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048346.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508048375.078, "dur": 21.024,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048375.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508048396.230, "dur": 1.952,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048396.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048398.309, "dur": 4.801,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048398.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048403.238, "dur": 30.816,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048403.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508048434.182, "dur": 1.984,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048434.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508048436.293, "dur": 3.361,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048436.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508048439.782, "dur": 7.808,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048439.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508048447.686, "dur": 8.992,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048447.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508048456.806, "dur": 41.855,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048456.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508048498.758, "dur": 8.352,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048498.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048507.238, "dur": 17.504,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048507.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508048524.870, "dur": 12.768,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048524.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048537.766, "dur": 50.592,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048537.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508048588.454, "dur": 13.504,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048588.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048602.054, "dur": 28.864,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048602.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508048631.046, "dur": 14.208,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048631.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508048645.382, "dur": 1.728,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048645.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048647.238, "dur": 3.840,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048647.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048651.206, "dur": 29.600,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048651.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508048680.934, "dur": 2.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048680.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508048683.078, "dur": 3.392,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048683.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508048686.598, "dur": 6.816,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048686.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508048693.542, "dur": 9.216,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048693.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508048702.854, "dur": 42.368,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048702.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508048745.350, "dur": 8.640,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048745.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048754.118, "dur": 17.216,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048754.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508048771.462, "dur": 15.744,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048771.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048787.334, "dur": 50.944,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048787.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508048838.406, "dur": 13.248,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048838.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048851.750, "dur": 28.800,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048851.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508048880.678, "dur": 21.696,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048880.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508048902.502, "dur": 1.920,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048902.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048904.582, "dur": 5.728,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048904.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508048910.438, "dur": 29.984,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048910.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508048940.550, "dur": 2.080,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048940.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508048942.758, "dur": 3.232,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048942.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508048946.086, "dur": 6.688,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048946.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508048952.870, "dur": 8.896,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048952.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508048961.894, "dur": 41.184,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508048961.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508049003.206, "dur": 8.416,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049003.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049011.750, "dur": 17.504,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049011.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508049029.382, "dur": 13.088,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049029.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049042.630, "dur": 50.080,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049042.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508049092.838, "dur": 13.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049092.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049106.246, "dur": 29.344,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049106.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508049135.718, "dur": 14.432,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049135.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508049150.278, "dur": 3.072,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049150.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049153.478, "dur": 3.264,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049153.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049156.870, "dur": 29.440,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049156.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508049186.406, "dur": 2.048,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049186.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508049188.550, "dur": 2.976,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049188.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508049191.622, "dur": 6.592,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049191.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508049198.342, "dur": 9.056,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049198.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508049207.526, "dur": 41.760,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049207.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508049249.414, "dur": 8.640,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049249.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049258.182, "dur": 16.672,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049258.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508049274.950, "dur": 15.872,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049274.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049290.982, "dur": 51.040,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049290.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508049342.182, "dur": 13.408,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049342.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049355.718, "dur": 28.768,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049355.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508049384.614, "dur": 21.664,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049384.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508049406.406, "dur": 2.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049406.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049408.550, "dur": 5.024,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049408.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049413.702, "dur": 30.400,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049413.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508049444.262, "dur": 2.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049444.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508049446.406, "dur": 3.360,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049446.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508049449.894, "dur": 6.752,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049449.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508049456.742, "dur": 8.800,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049456.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508049465.638, "dur": 43.904,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049465.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508049509.670, "dur": 8.704,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049509.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049518.502, "dur": 17.568,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049518.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508049536.198, "dur": 12.672,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049536.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049548.998, "dur": 49.504,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049548.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508049598.662, "dur": 13.440,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049598.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049612.230, "dur": 28.096,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049612.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508049640.454, "dur": 14.144,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049640.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508049654.726, "dur": 1.760,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049654.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049656.614, "dur": 3.776,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049656.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049660.518, "dur": 29.664,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049660.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508049690.278, "dur": 2.016,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049690.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508049692.422, "dur": 3.360,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049692.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508049695.910, "dur": 6.880,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049695.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508049702.886, "dur": 8.960,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049702.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508049711.974, "dur": 43.904,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049711.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508049756.006, "dur": 8.384,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049756.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049764.550, "dur": 17.184,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049764.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508049781.862, "dur": 16.128,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049781.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049798.118, "dur": 50.240,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049798.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508049848.486, "dur": 13.280,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049848.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049861.894, "dur": 28.032,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049861.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290508049890.022, "dur": 13.088,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049890.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290508049903.238, "dur": 6.176,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049903.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508049909.510, "dur": 147.456,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508049909.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508050056.870, "dur": 145.408,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050056.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508050202.150, "dur": 220.800,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050202.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290508050421.830, "dur": 21.920,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050421.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290508050443.686, "dur": 300.576,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050443.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508050747.366, "dur": 0.480,
    "args": {
      "External id": 163,      "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050747.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508050748.870, "dur": 139.008,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050748.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508050890.726, "dur": 1.152,
    "args": {
      "External id": 163,      "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "bytes": 576, "memory bandwidth (GB/s)": 0.5
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050890.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508050894.502, "dur": 142.656,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508050894.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508051037.062, "dur": 186.528,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051037.062,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290508051222.694, "dur": 21.504,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051222.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290508051244.134, "dur": 301.344,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051244.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508051547.206, "dur": 0.480,
    "args": {
      "External id": 163,      "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051547.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508051548.678, "dur": 138.208,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051548.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508051688.422, "dur": 1.280,
    "args": {
      "External id": 163,      "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "bytes": 576, "memory bandwidth (GB/s)": 0.45
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051688.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508051691.270, "dur": 139.840,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051691.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290508051831.430, "dur": 3.392,
    "args": {
      "External id": 163,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506990,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506990, "pid": 0, "tid": 7, "ts": 1290508051831.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290507998518.471, "dur": 369.473,
    "args": {
      "External id": 163,      "cbid": 311, "correlation": 506990
    }
  },
  {
    "ph": "s", "id": 506990, "pid": 14565, "tid": 14565, "ts": 1290507998518.471,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290508051836.742, "dur": 1.504,
    "args": {
      "External id": 180,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507003,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507003, "pid": 0, "tid": 7, "ts": 1290508051836.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290507999595.960, "dur": 12.952,
    "args": {
      "External id": 180,      "cbid": 211, "correlation": 507003
    }
  },
  {
    "ph": "s", "id": 507003, "pid": 14565, "tid": 14565, "ts": 1290507999595.960,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290508062292.808, "dur": 3.936,
    "args": {
      "External id": 188,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507826,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507826, "pid": 0, "tid": 7, "ts": 1290508062292.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290508007808.065, "dur": 12.639,
    "args": {
      "External id": 188,      "cbid": 211, "correlation": 507826
    }
  },
  {
    "ph": "s", "id": 507826, "pid": 14565, "tid": 14565, "ts": 1290508007808.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290508007827.578, "dur": 0.625,
    "args": {
      "External id": 186,      "cbid": 317, "correlation": 507834
    }
  },
  {
    "ph": "f", "id": 507834, "pid": 14565, "tid": 14565, "ts": 1290508007827.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508062300.648, "dur": 0.832,
    "args": {
      "External id": 186,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507836,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507836, "pid": 0, "tid": 7, "ts": 1290508062300.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508062301.640, "dur": 0.801,
    "args": {
      "External id": 186,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507836,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507836, "pid": 0, "tid": 7, "ts": 1290508062301.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290508062302.568, "dur": 2.080,
    "args": {
      "External id": 186,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507836,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507836, "pid": 0, "tid": 7, "ts": 1290508062302.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290508062304.776, "dur": 7.008,
    "args": {
      "External id": 186,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507836,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507836, "pid": 0, "tid": 7, "ts": 1290508062304.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290508062311.880, "dur": 11.200,
    "args": {
      "External id": 186,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507836,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507836, "pid": 0, "tid": 7, "ts": 1290508062311.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290508007828.904, "dur": 13.565,
    "args": {
      "External id": 186,      "cbid": 311, "correlation": 507836
    }
  },
  {
    "ph": "s", "id": 507836, "pid": 14565, "tid": 14565, "ts": 1290508007828.904,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508062325.512, "dur": 1.344,
    "args": {
      "External id": 196,      "device": 0, "context": 1,      "stream": 7, "correlation": 507841,      "bytes": 32, "memory bandwidth (GB/s)": 0.023809523809523808
    }
  },
  {
    "ph": "f", "id": 507841, "pid": 0, "tid": 7, "ts": 1290508062325.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508008250.601, "dur": 17.296,
    "args": {
      "External id": 196,      "cbid": 41, "correlation": 507841
    }
  },
  {
    "ph": "s", "id": 507841, "pid": 14565, "tid": 14565, "ts": 1290508008250.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508062327.272, "dur": 1.248,
    "args": {
      "External id": 197,      "device": 0, "context": 1,      "stream": 7, "correlation": 507846,      "bytes": 256, "memory bandwidth (GB/s)": 0.20512820512820512
    }
  },
  {
    "ph": "f", "id": 507846, "pid": 0, "tid": 7, "ts": 1290508062327.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508008273.442, "dur": 4.789,
    "args": {
      "External id": 197,      "cbid": 41, "correlation": 507846
    }
  },
  {
    "ph": "s", "id": 507846, "pid": 14565, "tid": 14565, "ts": 1290508008273.442,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508062328.936, "dur": 1.376,
    "args": {
      "External id": 198,      "device": 0, "context": 1,      "stream": 7, "correlation": 507851,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 507851, "pid": 0, "tid": 7, "ts": 1290508062328.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508008281.532, "dur": 3.948,
    "args": {
      "External id": 198,      "cbid": 41, "correlation": 507851
    }
  },
  {
    "ph": "s", "id": 507851, "pid": 14565, "tid": 14565, "ts": 1290508008281.532,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508062330.761, "dur": 1.311,
    "args": {
      "External id": 199,      "device": 0, "context": 1,      "stream": 7, "correlation": 507856,      "bytes": 256, "memory bandwidth (GB/s)": 0.1952707856598017
    }
  },
  {
    "ph": "f", "id": 507856, "pid": 0, "tid": 7, "ts": 1290508062330.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508008288.742, "dur": 4.422,
    "args": {
      "External id": 199,      "cbid": 41, "correlation": 507856
    }
  },
  {
    "ph": "s", "id": 507856, "pid": 14565, "tid": 14565, "ts": 1290508008288.742,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508062332.488, "dur": 2.080,
    "args": {
      "External id": 200,      "device": 0, "context": 1,      "stream": 7, "correlation": 507861,      "bytes": 65536, "memory bandwidth (GB/s)": 31.50769230769231
    }
  },
  {
    "ph": "f", "id": 507861, "pid": 0, "tid": 7, "ts": 1290508062332.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508008296.247, "dur": 3.904,
    "args": {
      "External id": 200,      "cbid": 41, "correlation": 507861
    }
  },
  {
    "ph": "s", "id": 507861, "pid": 14565, "tid": 14565, "ts": 1290508008296.247,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290508008306.932, "dur": 0.589,
    "args": {
      "External id": 193,      "cbid": 317, "correlation": 507871
    }
  },
  {
    "ph": "f", "id": 507871, "pid": 14565, "tid": 14565, "ts": 1290508008306.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062339.592, "dur": 1.504,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062339.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062341.192, "dur": 1.313,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062341.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062342.600, "dur": 1.312,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062342.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062344.041, "dur": 1.183,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062344.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062345.481, "dur": 1.439,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062345.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062347.048, "dur": 1.184,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062347.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062348.457, "dur": 1.439,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062348.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062350.024, "dur": 1.152,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062350.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062351.336, "dur": 1.440,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062351.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062352.904, "dur": 1.152,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062352.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062354.185, "dur": 1.503,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062354.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062355.785, "dur": 1.183,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062355.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062357.096, "dur": 1.376,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062357.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062358.568, "dur": 1.313,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062358.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062359.977, "dur": 1.247,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062359.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508062361.352, "dur": 1.248,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062361.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508062362.760, "dur": 2.273,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062362.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508062365.193, "dur": 1.856,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062365.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508062367.209, "dur": 1.791,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062367.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508062369.160, "dur": 1.792,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062369.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062371.081, "dur": 2.207,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062371.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062373.417, "dur": 1.951,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062373.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062375.464, "dur": 2.049,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062375.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062377.608, "dur": 2.016,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062377.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062379.720, "dur": 1.889,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062379.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062381.736, "dur": 2.369,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062381.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062384.232, "dur": 1.984,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062384.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062386.312, "dur": 2.080,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062386.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062388.552, "dur": 1.921,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062388.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062390.568, "dur": 2.208,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062390.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062392.873, "dur": 1.823,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062392.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062395.016, "dur": 1.792,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062395.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062396.936, "dur": 1.729,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062396.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062398.792, "dur": 1.792,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062398.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062400.712, "dur": 2.432,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062400.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508062403.273, "dur": 1.983,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062403.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062405.384, "dur": 3.136,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062405.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062408.616, "dur": 3.329,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062408.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062412.072, "dur": 2.240,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062412.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062414.441, "dur": 2.783,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062414.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062417.353, "dur": 2.687,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062417.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062420.169, "dur": 2.111,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062420.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062422.409, "dur": 2.559,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062422.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062425.096, "dur": 2.912,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062425.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062428.137, "dur": 3.167,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062428.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062431.433, "dur": 2.496,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062431.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062434.056, "dur": 2.784,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062434.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062436.968, "dur": 3.168,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062436.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062440.264, "dur": 2.688,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062440.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062443.048, "dur": 2.528,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062443.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062445.736, "dur": 2.753,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062445.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062448.616, "dur": 3.040,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062448.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062451.752, "dur": 2.304,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062451.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062454.184, "dur": 2.368,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062454.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062456.680, "dur": 3.009,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062456.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062459.816, "dur": 2.753,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062459.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062462.696, "dur": 2.336,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062462.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062465.128, "dur": 3.073,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062465.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062468.328, "dur": 2.656,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062468.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062471.080, "dur": 2.689,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062471.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062473.896, "dur": 2.433,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062473.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062476.456, "dur": 2.752,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062476.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062479.305, "dur": 2.463,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062479.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062481.897, "dur": 3.136,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062481.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062485.160, "dur": 3.104,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062485.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062488.393, "dur": 2.719,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062488.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062491.272, "dur": 2.593,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062491.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508062494.025, "dur": 2.847,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062494.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062497.000, "dur": 4.224,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062497.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062501.353, "dur": 4.959,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062501.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062506.441, "dur": 3.871,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062506.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062510.408, "dur": 4.576,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062510.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062515.080, "dur": 3.968,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062515.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062519.145, "dur": 4.096,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062519.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062523.368, "dur": 4.608,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062523.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062528.104, "dur": 4.128,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062528.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062532.329, "dur": 4.543,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062532.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062536.968, "dur": 3.968,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062536.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062541.064, "dur": 4.705,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062541.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062545.896, "dur": 4.225,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062545.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062550.248, "dur": 4.608,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062550.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062554.984, "dur": 4.416,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062554.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062559.496, "dur": 4.448,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062559.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508062564.041, "dur": 4.415,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062564.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290508062568.553, "dur": 7.519,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062568.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290508062576.232, "dur": 7.201,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062576.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062583.560, "dur": 12.449,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062583.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062596.104, "dur": 12.545,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062596.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062608.777, "dur": 12.351,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062608.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062621.224, "dur": 12.609,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062621.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290508062633.960, "dur": 16.033,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062633.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290508062650.120, "dur": 1.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062650.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062651.529, "dur": 3.007,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062651.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508062654.696, "dur": 21.953,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062654.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062676.777, "dur": 29.823,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062676.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508062706.729, "dur": 2.080,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062706.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508062708.905, "dur": 3.552,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062708.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508062712.584, "dur": 7.201,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062712.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508062719.913, "dur": 9.375,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062719.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508062729.384, "dur": 43.360,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062729.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508062772.873, "dur": 8.831,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062772.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062781.801, "dur": 17.823,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062781.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290508062799.784, "dur": 8.544,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062799.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062808.457, "dur": 49.312,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062808.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508062857.929, "dur": 13.472,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062857.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062871.561, "dur": 29.311,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062871.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508062901.032, "dur": 21.441,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062901.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508062922.600, "dur": 9.568,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062922.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508062931.081, "dur": 27.392,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062931.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508062958.601, "dur": 1.663,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062958.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062960.393, "dur": 4.000,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062960.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062964.489, "dur": 29.407,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062964.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508062994.025, "dur": 2.047,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062994.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508062996.201, "dur": 3.103,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062996.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508062999.401, "dur": 7.328,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508062999.401,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508063006.856, "dur": 8.897,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063006.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508063015.849, "dur": 41.344,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063015.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508063057.320, "dur": 8.737,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063057.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063066.217, "dur": 17.600,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063066.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508063083.944, "dur": 17.889,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063083.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063101.993, "dur": 50.527,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063101.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508063152.649, "dur": 13.248,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063152.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063166.024, "dur": 29.633,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063166.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508063195.785, "dur": 14.015,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063195.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508063209.929, "dur": 1.664,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063209.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063211.721, "dur": 3.807,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063211.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063215.625, "dur": 29.856,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063215.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508063245.577, "dur": 2.080,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063245.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508063247.785, "dur": 3.424,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063247.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508063251.336, "dur": 6.241,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063251.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508063257.705, "dur": 9.216,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063257.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508063267.017, "dur": 42.495,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063267.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508063309.641, "dur": 8.800,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063309.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063318.569, "dur": 17.023,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063318.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508063335.721, "dur": 15.904,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063335.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063351.753, "dur": 50.784,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063351.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508063402.665, "dur": 13.184,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063402.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063415.977, "dur": 29.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063415.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508063445.352, "dur": 23.233,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063445.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508063468.713, "dur": 3.232,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063468.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063472.041, "dur": 5.503,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063472.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063477.641, "dur": 29.536,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063477.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508063507.337, "dur": 2.016,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063507.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508063509.513, "dur": 3.424,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063509.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508063513.097, "dur": 6.688,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063513.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508063519.881, "dur": 9.184,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063519.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508063529.193, "dur": 42.400,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063529.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508063571.753, "dur": 8.512,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063571.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063580.425, "dur": 16.192,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063580.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508063596.745, "dur": 13.024,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063596.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063609.865, "dur": 49.632,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063609.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508063659.625, "dur": 13.376,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063659.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063673.097, "dur": 27.936,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063673.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508063701.161, "dur": 14.496,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063701.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508063715.785, "dur": 2.560,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063715.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063718.473, "dur": 3.488,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063718.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063722.089, "dur": 29.376,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063722.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508063751.561, "dur": 2.048,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063751.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508063753.737, "dur": 3.520,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063753.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508063757.417, "dur": 6.528,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063757.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508063764.041, "dur": 9.088,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063764.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508063773.225, "dur": 43.008,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063773.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508063816.393, "dur": 8.640,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063816.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063825.161, "dur": 17.152,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063825.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508063842.441, "dur": 15.936,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063842.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063858.505, "dur": 50.208,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063858.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508063908.841, "dur": 13.312,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063908.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063922.281, "dur": 29.376,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063922.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508063951.753, "dur": 24.096,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063951.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508063974.793, "dur": 2.080,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063974.793,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063977.001, "dur": 4.960,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063977.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508063982.089, "dur": 30.176,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508063982.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508064012.393, "dur": 2.016,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064012.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508064014.505, "dur": 3.392,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064014.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508064018.025, "dur": 7.040,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064018.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508064025.161, "dur": 8.800,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064025.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508064034.057, "dur": 42.464,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064034.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508064076.649, "dur": 8.512,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064076.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064085.289, "dur": 17.600,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064085.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508064103.017, "dur": 12.800,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064103.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064115.945, "dur": 49.824,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064115.945,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508064165.897, "dur": 13.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064165.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064179.305, "dur": 28.576,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064179.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508064208.009, "dur": 14.304,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064208.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508064222.441, "dur": 1.824,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064222.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064224.393, "dur": 3.840,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064224.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064228.329, "dur": 29.664,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064228.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508064258.121, "dur": 2.048,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064258.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508064260.265, "dur": 3.424,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064260.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508064263.817, "dur": 7.360,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064263.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508064271.305, "dur": 9.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064271.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508064280.713, "dur": 41.792,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064280.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508064322.633, "dur": 8.512,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064322.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064331.273, "dur": 16.640,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064331.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508064348.041, "dur": 15.904,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064348.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064364.073, "dur": 50.720,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064364.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508064414.953, "dur": 13.184,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064414.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064428.297, "dur": 27.584,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064428.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508064456.009, "dur": 22.272,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064456.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508064478.409, "dur": 2.432,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064478.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064480.937, "dur": 4.896,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064480.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064485.929, "dur": 31.264,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064485.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508064517.321, "dur": 2.048,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064517.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508064519.465, "dur": 3.424,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064519.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508064523.049, "dur": 6.720,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064523.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508064529.865, "dur": 9.248,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064529.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508064539.209, "dur": 43.232,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064539.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508064582.569, "dur": 8.416,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064582.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064591.145, "dur": 17.792,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064591.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508064609.065, "dur": 12.768,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064609.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064621.961, "dur": 50.656,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064621.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508064672.745, "dur": 13.376,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064672.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064686.249, "dur": 28.288,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064686.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508064714.665, "dur": 22.528,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064714.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508064737.321, "dur": 8.608,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064737.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290508064746.089, "dur": 27.552,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064746.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508064773.737, "dur": 2.176,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064773.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064776.041, "dur": 3.264,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064776.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064779.433, "dur": 29.952,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064779.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508064809.513, "dur": 2.048,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064809.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508064811.689, "dur": 3.424,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064811.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508064815.273, "dur": 6.752,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064815.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508064822.121, "dur": 9.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064822.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508064831.529, "dur": 43.296,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064831.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508064874.953, "dur": 8.832,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064874.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064883.913, "dur": 17.632,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064883.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508064901.673, "dur": 15.360,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064901.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064917.161, "dur": 51.072,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064917.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508064968.361, "dur": 13.248,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064968.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508064981.705, "dur": 28.160,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508064981.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508065009.993, "dur": 14.432,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065009.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508065024.553, "dur": 1.888,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065024.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065026.569, "dur": 3.968,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065026.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065030.665, "dur": 29.568,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065030.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508065060.329, "dur": 2.080,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065060.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508065062.505, "dur": 3.424,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065062.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508065066.057, "dur": 6.624,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065066.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508065072.809, "dur": 9.312,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065072.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508065082.217, "dur": 42.912,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065082.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508065125.257, "dur": 8.704,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065125.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065134.121, "dur": 16.992,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065134.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508065151.273, "dur": 15.840,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065151.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065167.209, "dur": 50.816,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065167.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508065218.153, "dur": 13.248,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065218.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065231.497, "dur": 28.288,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065231.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508065259.913, "dur": 21.184,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065259.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508065281.225, "dur": 3.040,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065281.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065284.393, "dur": 4.960,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065284.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065289.449, "dur": 29.920,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065289.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508065319.497, "dur": 1.984,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065319.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508065321.641, "dur": 3.360,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065321.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508065325.161, "dur": 6.464,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065325.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508065331.753, "dur": 8.992,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065331.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508065340.873, "dur": 42.848,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065340.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508065383.849, "dur": 8.384,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065383.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065392.329, "dur": 17.504,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065392.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508065409.961, "dur": 12.704,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065409.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065422.793, "dur": 51.328,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065422.793,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508065474.249, "dur": 13.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065474.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065487.657, "dur": 29.344,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065487.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508065517.129, "dur": 14.336,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065517.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508065531.593, "dur": 1.696,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065531.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065533.417, "dur": 3.808,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065533.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065537.353, "dur": 29.472,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065537.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508065566.953, "dur": 2.016,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065566.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508065569.097, "dur": 3.392,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065569.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508065572.617, "dur": 7.488,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065572.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508065580.233, "dur": 9.056,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065580.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508065589.385, "dur": 41.248,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065589.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508065630.761, "dur": 8.608,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065630.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065639.497, "dur": 17.120,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065639.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508065656.745, "dur": 15.808,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065656.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065672.681, "dur": 51.168,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065672.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508065723.977, "dur": 13.216,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065723.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065737.321, "dur": 28.704,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065737.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508065766.153, "dur": 22.016,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065766.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508065788.297, "dur": 2.912,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065788.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065791.337, "dur": 5.760,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065791.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065797.225, "dur": 29.760,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065797.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508065827.113, "dur": 2.080,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065827.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508065829.321, "dur": 3.296,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065829.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508065832.745, "dur": 6.656,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065832.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508065839.529, "dur": 8.928,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065839.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508065848.585, "dur": 40.928,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065848.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508065889.609, "dur": 8.448,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065889.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065898.153, "dur": 17.504,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065898.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508065915.817, "dur": 13.344,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065915.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065929.289, "dur": 49.632,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065929.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508065979.049, "dur": 13.408,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065979.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508065992.585, "dur": 29.408,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508065992.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508066022.121, "dur": 14.400,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066022.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508066036.617, "dur": 1.792,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066036.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066038.569, "dur": 4.256,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066038.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066042.953, "dur": 29.536,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066042.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508066072.617, "dur": 2.048,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066072.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508066074.761, "dur": 3.392,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066074.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508066078.281, "dur": 6.592,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066078.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508066085.001, "dur": 9.152,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066085.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508066094.249, "dur": 42.880,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066094.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508066137.257, "dur": 8.736,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066137.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066146.121, "dur": 16.672,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066146.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508066162.889, "dur": 15.872,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066162.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066178.921, "dur": 51.040,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066178.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508066230.121, "dur": 13.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066230.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066243.561, "dur": 28.704,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066243.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508066272.393, "dur": 21.984,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066272.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508066294.505, "dur": 1.888,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066294.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066296.521, "dur": 4.896,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066296.521,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066301.577, "dur": 30.336,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066301.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508066332.041, "dur": 2.048,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066332.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508066334.185, "dur": 3.488,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066334.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508066337.769, "dur": 6.784,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066337.769,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508066344.649, "dur": 8.960,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066344.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508066353.705, "dur": 44.000,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066353.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508066397.833, "dur": 8.768,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066397.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066406.729, "dur": 17.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066406.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508066424.137, "dur": 12.704,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066424.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066436.969, "dur": 49.440,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066436.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508066486.569, "dur": 13.280,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066486.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066499.977, "dur": 28.352,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066499.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508066528.457, "dur": 14.176,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066528.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508066542.761, "dur": 1.760,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066542.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066544.649, "dur": 3.744,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066544.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066548.521, "dur": 29.856,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066548.521,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508066578.505, "dur": 2.048,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066578.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508066580.681, "dur": 3.392,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066580.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508066584.201, "dur": 6.784,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066584.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508066591.113, "dur": 9.216,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066591.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508066600.457, "dur": 43.968,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066600.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508066644.553, "dur": 8.448,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066644.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066653.161, "dur": 17.408,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066653.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508066670.697, "dur": 16.192,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066670.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066687.017, "dur": 50.880,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066687.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508066738.057, "dur": 13.152,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066738.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066751.337, "dur": 28.032,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066751.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290508066779.497, "dur": 12.928,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066779.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290508066792.553, "dur": 6.208,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066792.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066798.889, "dur": 149.824,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066798.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508066948.585, "dur": 147.072,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508066948.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508067095.561, "dur": 223.360,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067095.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290508067317.833, "dur": 19.617,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067317.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290508067337.385, "dur": 303.137,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067337.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508067643.338, "dur": 0.480,
    "args": {
      "External id": 193,      "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067643.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508067644.841, "dur": 137.953,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067644.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508067784.745, "dur": 1.152,
    "args": {
      "External id": 193,      "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "bytes": 576, "memory bandwidth (GB/s)": 0.5
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067784.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508067788.490, "dur": 141.119,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067788.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508067929.514, "dur": 185.856,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508067929.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290508068114.762, "dur": 20.800,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508068114.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290508068135.465, "dur": 299.489,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508068135.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508068437.034, "dur": 0.480,
    "args": {
      "External id": 193,      "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508068437.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508068438.538, "dur": 138.528,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508068438.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508068580.426, "dur": 1.312,
    "args": {
      "External id": 193,      "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "bytes": 576, "memory bandwidth (GB/s)": 0.43902439024390244
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508068580.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508068583.338, "dur": 140.672,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508068583.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290508068724.330, "dur": 3.328,
    "args": {
      "External id": 193,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507873,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507873, "pid": 0, "tid": 7, "ts": 1290508068724.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290508008308.105, "dur": 348.739,
    "args": {
      "External id": 193,      "cbid": 311, "correlation": 507873
    }
  },
  {
    "ph": "s", "id": 507873, "pid": 14565, "tid": 14565, "ts": 1290508008308.105,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290508068730.698, "dur": 1.664,
    "args": {
      "External id": 210,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507886,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507886, "pid": 0, "tid": 7, "ts": 1290508068730.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290508009405.870, "dur": 17.268,
    "args": {
      "External id": 210,      "cbid": 211, "correlation": 507886
    }
  },
  {
    "ph": "s", "id": 507886, "pid": 14565, "tid": 14565, "ts": 1290508009405.870,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<long, 2, 1, 1>, at::native::Copy<long, long>)", "pid": 0, "tid": 7,
    "ts": 1290508079189.740, "dur": 3.904,
    "args": {
      "External id": 218,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508709,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508709, "pid": 0, "tid": 7, "ts": 1290508079189.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290508013913.500, "dur": 17.362,
    "args": {
      "External id": 218,      "cbid": 211, "correlation": 508709
    }
  },
  {
    "ph": "s", "id": 508709, "pid": 14565, "tid": 14565, "ts": 1290508013913.500,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290508013939.320, "dur": 0.736,
    "args": {
      "External id": 216,      "cbid": 317, "correlation": 508717
    }
  },
  {
    "ph": "f", "id": 508717, "pid": 14565, "tid": 14565, "ts": 1290508013939.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508079197.836, "dur": 0.832,
    "args": {
      "External id": 216,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508719,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508719, "pid": 0, "tid": 7, "ts": 1290508079197.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_zeros_0", "pid": 0, "tid": 7,
    "ts": 1290508079198.796, "dur": 0.832,
    "args": {
      "External id": 216,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508719,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.005319,      "grid": [1, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508719, "pid": 0, "tid": 7, "ts": 1290508079198.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_bitwise_xor_constant_pad_nd_index_mul_select_slice_1", "pid": 0, "tid": 7,
    "ts": 1290508079199.756, "dur": 2.112,
    "args": {
      "External id": 216,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508719,      "registers per thread": 32,      "shared memory": 0,      "blocks per SM": 0.340426,      "warps per SM": 1.361702,      "grid": [64, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508719, "pid": 0, "tid": 7, "ts": 1290508079199.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_bitwise_xor_constant_pad_nd_copy_index_mul_remainder_select_slice_unsqueeze_2", "pid": 0, "tid": 7,
    "ts": 1290508079202.028, "dur": 6.848,
    "args": {
      "External id": 216,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508719,      "registers per thread": 114,      "shared memory": 4096,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508719, "pid": 0, "tid": 7, "ts": 1290508079202.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_3", "pid": 0, "tid": 7,
    "ts": 1290508079209.004, "dur": 10.496,
    "args": {
      "External id": 216,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508719,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508719, "pid": 0, "tid": 7, "ts": 1290508079209.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290508013940.784, "dur": 15.012,
    "args": {
      "External id": 216,      "cbid": 311, "correlation": 508719
    }
  },
  {
    "ph": "s", "id": 508719, "pid": 14565, "tid": 14565, "ts": 1290508013940.784,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508079220.620, "dur": 1.888,
    "args": {
      "External id": 226,      "device": 0, "context": 1,      "stream": 7, "correlation": 508724,      "bytes": 32, "memory bandwidth (GB/s)": 0.01694915254237288
    }
  },
  {
    "ph": "f", "id": 508724, "pid": 0, "tid": 7, "ts": 1290508079220.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508014445.841, "dur": 19.724,
    "args": {
      "External id": 226,      "cbid": 41, "correlation": 508724
    }
  },
  {
    "ph": "s", "id": 508724, "pid": 14565, "tid": 14565, "ts": 1290508014445.841,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508079222.924, "dur": 1.280,
    "args": {
      "External id": 227,      "device": 0, "context": 1,      "stream": 7, "correlation": 508729,      "bytes": 256, "memory bandwidth (GB/s)": 0.2
    }
  },
  {
    "ph": "f", "id": 508729, "pid": 0, "tid": 7, "ts": 1290508079222.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508014471.321, "dur": 4.939,
    "args": {
      "External id": 227,      "cbid": 41, "correlation": 508729
    }
  },
  {
    "ph": "s", "id": 508729, "pid": 14565, "tid": 14565, "ts": 1290508014471.321,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508079224.620, "dur": 1.376,
    "args": {
      "External id": 228,      "device": 0, "context": 1,      "stream": 7, "correlation": 508734,      "bytes": 32, "memory bandwidth (GB/s)": 0.023255813953488372
    }
  },
  {
    "ph": "f", "id": 508734, "pid": 0, "tid": 7, "ts": 1290508079224.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508014479.317, "dur": 4.346,
    "args": {
      "External id": 228,      "cbid": 41, "correlation": 508734
    }
  },
  {
    "ph": "s", "id": 508734, "pid": 14565, "tid": 14565, "ts": 1290508014479.317,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508079226.412, "dur": 1.312,
    "args": {
      "External id": 229,      "device": 0, "context": 1,      "stream": 7, "correlation": 508739,      "bytes": 256, "memory bandwidth (GB/s)": 0.1951219512195122
    }
  },
  {
    "ph": "f", "id": 508739, "pid": 0, "tid": 7, "ts": 1290508079226.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508014486.521, "dur": 4.020,
    "args": {
      "External id": 229,      "cbid": 41, "correlation": 508739
    }
  },
  {
    "ph": "s", "id": 508739, "pid": 14565, "tid": 14565, "ts": 1290508014486.521,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7,
    "ts": 1290508079228.140, "dur": 1.984,
    "args": {
      "External id": 230,      "device": 0, "context": 1,      "stream": 7, "correlation": 508744,      "bytes": 65536, "memory bandwidth (GB/s)": 33.03225806451613
    }
  },
  {
    "ph": "f", "id": 508744, "pid": 0, "tid": 7, "ts": 1290508079228.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 14565, "tid": 14565,
    "ts": 1290508014493.524, "dur": 3.549,
    "args": {
      "External id": 230,      "cbid": 41, "correlation": 508744
    }
  },
  {
    "ph": "s", "id": 508744, "pid": 14565, "tid": 14565, "ts": 1290508014493.524,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14565,
    "ts": 1290508014505.678, "dur": 0.720,
    "args": {
      "External id": 223,      "cbid": 317, "correlation": 508754
    }
  },
  {
    "ph": "f", "id": 508754, "pid": 14565, "tid": 14565, "ts": 1290508014505.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079235.596, "dur": 1.568,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079235.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079237.292, "dur": 1.312,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079237.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079238.700, "dur": 1.312,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079238.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079240.108, "dur": 1.248,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079240.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079241.580, "dur": 1.440,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079241.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079243.148, "dur": 1.184,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079243.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079244.556, "dur": 1.472,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079244.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079246.124, "dur": 1.184,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079246.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079247.660, "dur": 1.472,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079247.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079249.260, "dur": 1.536,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079249.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079250.892, "dur": 1.248,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079250.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079252.268, "dur": 1.184,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079252.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079253.804, "dur": 1.504,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079253.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079255.436, "dur": 1.312,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079255.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079256.876, "dur": 1.248,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079256.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_0", "pid": 0, "tid": 7,
    "ts": 1290508079258.252, "dur": 1.216,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.010638,      "warps per SM": 0.042553,      "grid": [2, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079258.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508079259.884, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079259.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508079262.092, "dur": 1.824,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079262.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508079264.044, "dur": 1.792,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079264.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_1", "pid": 0, "tid": 7,
    "ts": 1290508079266.060, "dur": 2.336,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079266.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079268.524, "dur": 2.688,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079268.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079271.340, "dur": 1.920,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079271.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079273.388, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079273.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079275.564, "dur": 2.432,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079275.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079278.092, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079278.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079280.268, "dur": 1.920,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079280.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079282.284, "dur": 2.112,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079282.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079284.620, "dur": 2.272,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079284.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079287.020, "dur": 2.304,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079287.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079289.644, "dur": 1.760,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079289.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079291.532, "dur": 1.600,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079291.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079293.260, "dur": 1.920,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079293.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079295.308, "dur": 1.600,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079295.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079297.036, "dur": 1.664,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079297.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079298.828, "dur": 2.112,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079298.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_2", "pid": 0, "tid": 7,
    "ts": 1290508079301.068, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 12.255320,      "grid": [288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079301.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079303.212, "dur": 3.264,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079303.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079306.572, "dur": 3.008,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079306.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079309.708, "dur": 2.176,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079309.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079312.012, "dur": 2.752,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079312.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079314.892, "dur": 3.040,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079314.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079318.060, "dur": 2.112,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079318.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079320.300, "dur": 2.496,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079320.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079322.924, "dur": 2.816,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079322.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079325.868, "dur": 2.816,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079325.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079328.780, "dur": 2.496,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079328.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079331.372, "dur": 2.720,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079331.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079334.220, "dur": 2.976,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079334.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079337.324, "dur": 2.720,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079337.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079340.140, "dur": 2.592,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079340.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079342.860, "dur": 3.232,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079342.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079346.220, "dur": 2.912,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079346.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079349.260, "dur": 2.272,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079349.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079351.660, "dur": 2.336,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079351.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079354.124, "dur": 2.528,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079354.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079356.748, "dur": 2.752,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079356.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079359.596, "dur": 2.336,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079359.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079362.060, "dur": 3.232,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079362.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079365.420, "dur": 2.560,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079365.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079368.108, "dur": 2.720,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079368.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079370.924, "dur": 2.752,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079370.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079373.772, "dur": 2.688,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079373.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079376.588, "dur": 2.304,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079376.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079379.020, "dur": 3.072,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079379.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079382.220, "dur": 2.752,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079382.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079385.132, "dur": 2.688,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079385.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079387.948, "dur": 2.592,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079387.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_t_3", "pid": 0, "tid": 7,
    "ts": 1290508079390.700, "dur": 2.912,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079390.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079393.708, "dur": 4.128,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079393.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079397.964, "dur": 4.832,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079397.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079402.924, "dur": 3.872,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079402.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079406.924, "dur": 4.864,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079406.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079411.884, "dur": 4.160,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079411.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079416.140, "dur": 4.224,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079416.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079420.524, "dur": 4.192,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079420.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079424.844, "dur": 4.128,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079424.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079429.100, "dur": 4.608,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079429.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079433.836, "dur": 3.968,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079433.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079437.932, "dur": 4.960,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079437.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079443.020, "dur": 4.160,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079443.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079447.276, "dur": 4.800,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079447.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079452.204, "dur": 4.416,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079452.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079456.716, "dur": 4.512,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079456.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_cat_4", "pid": 0, "tid": 7,
    "ts": 1290508079461.356, "dur": 4.448,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 26,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079461.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_5", "pid": 0, "tid": 7,
    "ts": 1290508079465.932, "dur": 7.488,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079465.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_embedding_view_6", "pid": 0, "tid": 7,
    "ts": 1290508079473.580, "dur": 7.104,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 24,      "shared memory": 0,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [2048, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079473.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079480.812, "dur": 12.448,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079480.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079493.388, "dur": 12.640,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079493.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079506.156, "dur": 12.320,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079506.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079518.604, "dur": 12.512,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079518.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy_7", "pid": 0, "tid": 7,
    "ts": 1290508079531.276, "dur": 16.000,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 65,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079531.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_8", "pid": 0, "tid": 7,
    "ts": 1290508079547.404, "dur": 1.280,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 28,      "shared memory": 0,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [1, 128, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079547.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079548.812, "dur": 2.976,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079548.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508079551.916, "dur": 22.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 18,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079551.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079574.092, "dur": 29.632,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079574.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508079603.852, "dur": 2.112,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079603.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508079606.060, "dur": 3.488,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079606.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508079609.676, "dur": 6.688,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079609.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508079616.492, "dur": 9.472,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079616.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508079626.092, "dur": 42.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079626.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508079668.236, "dur": 8.640,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079668.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079677.004, "dur": 17.856,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079677.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_16", "pid": 0, "tid": 7,
    "ts": 1290508079695.020, "dur": 8.512,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 70,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079695.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079703.692, "dur": 49.280,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079703.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508079753.132, "dur": 13.440,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079753.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079766.732, "dur": 29.664,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079766.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508079796.524, "dur": 21.920,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079796.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508079818.572, "dur": 9.664,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079818.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508079827.148, "dur": 27.744,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079827.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508079855.020, "dur": 2.272,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079855.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079857.388, "dur": 3.712,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079857.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079861.196, "dur": 30.432,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079861.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508079891.724, "dur": 2.080,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079891.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508079893.932, "dur": 3.552,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079893.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508079897.612, "dur": 6.688,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079897.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508079904.396, "dur": 9.248,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079904.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508079913.740, "dur": 42.560,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079913.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508079956.396, "dur": 8.992,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079956.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079965.516, "dur": 17.728,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079965.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508079983.372, "dur": 16.480,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079983.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079999.980, "dur": 52.864,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508079999.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508080053.004, "dur": 13.248,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080053.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080066.380, "dur": 29.216,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080066.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508080095.692, "dur": 14.112,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080095.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508080109.932, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080109.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080112.108, "dur": 3.744,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080112.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080115.948, "dur": 30.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080115.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508080146.060, "dur": 2.080,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080146.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508080148.268, "dur": 3.456,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080148.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508080151.852, "dur": 6.688,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080151.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508080158.668, "dur": 9.280,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080158.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508080168.076, "dur": 43.136,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080168.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508080211.340, "dur": 8.736,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080211.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080220.204, "dur": 17.088,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080220.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508080237.420, "dur": 15.808,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080237.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080253.356, "dur": 51.104,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080253.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508080304.588, "dur": 13.312,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080304.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080318.028, "dur": 29.984,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080318.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508080348.140, "dur": 24.256,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080348.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508080372.524, "dur": 1.984,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080372.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080374.604, "dur": 4.960,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080374.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080379.660, "dur": 29.760,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080379.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508080409.548, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080409.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508080411.724, "dur": 4.192,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080411.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508080416.044, "dur": 6.528,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080416.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508080422.668, "dur": 9.152,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080422.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508080431.948, "dur": 45.120,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080431.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508080477.196, "dur": 8.736,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080477.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080486.092, "dur": 16.256,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080486.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508080502.508, "dur": 13.120,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080502.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080515.756, "dur": 50.240,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080515.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508080566.092, "dur": 13.504,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080566.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080579.724, "dur": 28.352,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080579.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508080608.204, "dur": 14.592,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080608.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508080622.924, "dur": 1.824,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080622.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080624.876, "dur": 3.616,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080624.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080628.588, "dur": 29.696,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080628.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508080658.412, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080658.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508080660.620, "dur": 3.552,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080660.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508080664.300, "dur": 6.656,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080664.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508080671.052, "dur": 8.928,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080671.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508080680.076, "dur": 41.952,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080680.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508080722.156, "dur": 8.544,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080722.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080730.828, "dur": 17.184,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080730.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508080748.172, "dur": 15.872,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080748.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080764.172, "dur": 50.432,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080764.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508080814.732, "dur": 13.312,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080814.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080828.172, "dur": 29.024,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080828.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508080857.324, "dur": 23.936,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080857.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508080880.332, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080880.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080882.444, "dur": 4.192,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080882.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080886.764, "dur": 30.944,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080886.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508080917.804, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080917.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508080919.980, "dur": 3.424,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080919.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508080923.532, "dur": 6.752,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080923.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508080930.412, "dur": 8.928,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080930.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508080939.468, "dur": 43.008,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080939.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508080982.604, "dur": 8.512,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080982.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508080991.244, "dur": 17.408,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508080991.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508081008.780, "dur": 12.768,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081008.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081021.676, "dur": 49.984,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081021.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508081071.820, "dur": 13.280,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081071.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081085.228, "dur": 29.088,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081085.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508081114.444, "dur": 14.272,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081114.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508081128.844, "dur": 1.760,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081128.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081130.732, "dur": 4.064,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081130.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081134.892, "dur": 29.568,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081134.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508081164.588, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081164.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508081166.764, "dur": 3.424,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081166.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508081170.316, "dur": 6.816,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081170.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508081177.260, "dur": 9.184,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081177.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508081186.572, "dur": 42.976,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081186.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508081229.676, "dur": 8.576,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081229.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081238.380, "dur": 16.928,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081238.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508081255.436, "dur": 15.808,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081255.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081271.372, "dur": 50.272,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081271.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508081321.804, "dur": 13.312,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081321.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081335.276, "dur": 28.096,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081335.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508081363.500, "dur": 22.848,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081363.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508081386.476, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081386.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081388.620, "dur": 4.960,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081388.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081393.676, "dur": 31.232,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081393.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508081425.036, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081425.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508081427.180, "dur": 3.392,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081427.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508081430.700, "dur": 6.432,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081430.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508081437.260, "dur": 9.120,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081437.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508081446.476, "dur": 44.544,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081446.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508081491.148, "dur": 8.576,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081491.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081499.852, "dur": 17.856,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081499.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508081517.868, "dur": 12.672,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081517.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081530.668, "dur": 50.848,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081530.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508081581.644, "dur": 13.440,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081581.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081595.180, "dur": 28.160,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081595.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_clamp_min_div_mul_sigmoid_sign_sqrt_sum_unsqueeze_18", "pid": 0, "tid": 7,
    "ts": 1290508081623.468, "dur": 22.432,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081623.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_19", "pid": 0, "tid": 7,
    "ts": 1290508081646.028, "dur": 8.576,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081646.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__to_copy__unsafe_view_abs_add_addcmul_clamp_min_mul_select_sigmoid_sign_silu_slice_sqrt_unsqueeze_26", "pid": 0, "tid": 7,
    "ts": 1290508081654.732, "dur": 27.712,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 48,      "shared memory": 2048,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [4096, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 83
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081654.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508081682.572, "dur": 1.856,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081682.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081684.556, "dur": 3.296,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081684.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081687.980, "dur": 29.440,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081687.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508081717.580, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081717.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508081719.724, "dur": 3.456,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081719.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508081723.308, "dur": 6.528,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081723.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508081729.932, "dur": 9.024,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081729.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508081739.052, "dur": 44.928,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081739.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508081784.108, "dur": 8.768,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081784.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081793.004, "dur": 17.664,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081793.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508081810.796, "dur": 16.864,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081810.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081827.756, "dur": 50.464,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081827.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508081878.316, "dur": 13.216,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081878.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081891.628, "dur": 28.192,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081891.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508081919.980, "dur": 14.240,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081919.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508081934.348, "dur": 1.824,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081934.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081936.300, "dur": 3.872,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081936.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508081940.300, "dur": 29.760,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081940.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508081970.156, "dur": 2.080,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081970.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508081972.332, "dur": 3.392,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081972.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508081975.852, "dur": 6.688,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081975.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508081982.668, "dur": 9.184,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081982.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508081991.948, "dur": 42.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508081991.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508082034.092, "dur": 8.576,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082034.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082042.828, "dur": 17.024,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082042.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508082059.980, "dur": 15.936,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082059.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082076.044, "dur": 50.689,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082076.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508082126.860, "dur": 13.153,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082126.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082140.140, "dur": 28.064,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082140.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508082168.332, "dur": 21.568,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082168.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508082190.028, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082190.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082192.204, "dur": 4.768,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082192.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082197.100, "dur": 30.336,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082197.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508082227.565, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082227.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508082229.741, "dur": 3.327,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082229.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508082233.228, "dur": 7.008,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082233.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508082240.365, "dur": 9.279,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082240.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508082249.740, "dur": 42.912,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082249.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508082292.780, "dur": 8.353,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082292.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082301.228, "dur": 17.888,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082301.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508082319.212, "dur": 12.737,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082319.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082332.044, "dur": 51.104,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082332.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508082383.277, "dur": 13.407,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082383.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082396.813, "dur": 28.831,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082396.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508082425.773, "dur": 14.207,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082425.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508082440.109, "dur": 1.696,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082440.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082441.932, "dur": 3.840,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082441.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082445.868, "dur": 29.569,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082445.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508082475.564, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082475.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508082477.708, "dur": 3.361,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082477.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508082481.229, "dur": 7.392,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082481.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508082488.748, "dur": 9.504,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082488.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508082498.348, "dur": 43.168,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082498.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508082541.645, "dur": 8.607,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082541.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082550.412, "dur": 16.896,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082550.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508082567.436, "dur": 15.841,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082567.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082583.404, "dur": 50.752,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082583.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508082634.285, "dur": 13.184,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082634.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082647.596, "dur": 29.376,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082647.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508082677.101, "dur": 22.399,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082677.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508082699.629, "dur": 1.983,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082699.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082701.740, "dur": 4.864,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082701.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082706.732, "dur": 30.465,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082706.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508082737.324, "dur": 2.049,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082737.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508082739.500, "dur": 3.361,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082739.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508082742.989, "dur": 6.719,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082742.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508082749.804, "dur": 8.928,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082749.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508082758.861, "dur": 42.687,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082758.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508082801.677, "dur": 8.479,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082801.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082810.285, "dur": 17.792,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082810.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508082828.237, "dur": 13.024,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082828.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082841.388, "dur": 50.208,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082841.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508082891.725, "dur": 13.471,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082891.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082905.325, "dur": 29.407,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082905.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508082934.861, "dur": 14.496,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082934.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508082949.485, "dur": 1.792,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082949.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082951.437, "dur": 3.936,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082951.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508082955.500, "dur": 29.505,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082955.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508082985.100, "dur": 2.081,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082985.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508082987.276, "dur": 3.585,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082987.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508082990.988, "dur": 6.561,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082990.988,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508082997.645, "dur": 9.023,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508082997.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508083006.764, "dur": 40.224,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083006.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508083047.117, "dur": 8.608,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083047.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083055.821, "dur": 17.312,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083055.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508083073.228, "dur": 15.873,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083073.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083089.229, "dur": 50.080,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083089.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508083139.469, "dur": 13.343,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083139.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083152.941, "dur": 28.800,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083152.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_25", "pid": 0, "tid": 7,
    "ts": 1290508083181.868, "dur": 22.945,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083181.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508083204.941, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083204.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083207.085, "dur": 5.599,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083207.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083212.813, "dur": 29.344,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083212.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508083242.317, "dur": 2.048,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083242.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508083244.461, "dur": 4.351,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083244.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508083248.908, "dur": 6.785,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083248.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508083255.789, "dur": 9.056,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083255.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508083264.973, "dur": 43.616,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083264.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508083308.685, "dur": 8.608,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083308.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083317.453, "dur": 17.120,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083317.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_22", "pid": 0, "tid": 7,
    "ts": 1290508083334.701, "dur": 12.768,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083334.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083347.597, "dur": 49.728,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083347.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508083397.452, "dur": 13.249,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083397.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083410.829, "dur": 28.224,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083410.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_23", "pid": 0, "tid": 7,
    "ts": 1290508083439.149, "dur": 14.208,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083439.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_21", "pid": 0, "tid": 7,
    "ts": 1290508083453.453, "dur": 1.215,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083453.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083454.957, "dur": 4.640,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 78,      "shared memory": 33792,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083454.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083459.693, "dur": 29.983,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 226,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083459.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_10", "pid": 0, "tid": 7,
    "ts": 1290508083489.805, "dur": 2.016,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [6, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083489.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_split_with_sizes_transpose_view_11", "pid": 0, "tid": 7,
    "ts": 1290508083491.949, "dur": 3.392,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 22,      "shared memory": 512,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [12, 64, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083491.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_12", "pid": 0, "tid": 7,
    "ts": 1290508083495.501, "dur": 6.912,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 62,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083495.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy__unsafe_view_add_cat_mul_neg_slice_split_split_with_sizes_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508083502.509, "dur": 9.024,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 68,      "shared memory": 4096,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 58
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083502.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_split_with_sizes_transpose_view_14", "pid": 0, "tid": 7,
    "ts": 1290508083511.629, "dur": 43.231,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 119,      "shared memory": 20480,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [16, 8, 12],      "block": [128, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083511.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_sigmoid_slice_split_with_sizes_sub_sum_tanh_transpose_unsqueeze_view_15", "pid": 0, "tid": 7,
    "ts": 1290508083554.989, "dur": 8.576,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 168,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 8.170213,      "grid": [768, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083554.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083563.693, "dur": 17.088,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083563.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__to_copy__unsafe_view_add_24", "pid": 0, "tid": 7,
    "ts": 1290508083580.909, "dur": 15.872,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 80,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083580.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083596.941, "dur": 50.848,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [256, 2, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083596.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_mul_silu_split_17", "pid": 0, "tid": 7,
    "ts": 1290508083647.917, "dur": 13.280,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 522.893616,      "grid": [12288, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083647.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083661.293, "dur": 28.000,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083661.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__unsafe_view_add_27", "pid": 0, "tid": 7,
    "ts": 1290508083689.389, "dur": 12.960,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 85,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083689.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__to_copy_slice_view_28", "pid": 0, "tid": 7,
    "ts": 1290508083702.445, "dur": 6.464,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 36,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083702.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083709.005, "dur": 147.520,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083709.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_tn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508083856.429, "dur": 145.600,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 8.170213,      "warps per SM": 65.361702,      "grid": [128, 12, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508083856.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508084001.901, "dur": 223.616,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084001.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_30", "pid": 0, "tid": 7,
    "ts": 1290508084224.653, "dur": 19.360,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084224.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_31", "pid": 0, "tid": 7,
    "ts": 1290508084243.917, "dur": 300.864,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084243.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508084547.501, "dur": 0.512,
    "args": {
      "External id": 223,      "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "bytes": 192, "memory bandwidth (GB/s)": 0.375
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084547.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508084549.037, "dur": 137.568,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084549.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508084687.981, "dur": 1.056,
    "args": {
      "External id": 223,      "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "bytes": 576, "memory bandwidth (GB/s)": 0.5454545454545454
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084687.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508084691.629, "dur": 140.800,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084691.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_logsumexp_ne_neg_scatter_add_slice_sub_unsqueeze_view_where_zeros_like_29", "pid": 0, "tid": 7,
    "ts": 1290508084832.333, "dur": 185.504,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 43,      "shared memory": 32768,      "blocks per SM": 21.787233,      "warps per SM": 697.191467,      "grid": [4096, 1, 1],      "block": [1024, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508084832.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_exp_gather_logsumexp_mul_ne_neg_scatter_add_slice_sub_sum_unsqueeze_view_where_zeros_like_32", "pid": 0, "tid": 7,
    "ts": 1290508085016.813, "dur": 20.512,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 54,      "shared memory": 16384,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508085016.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mul_ne_slice_t_unsqueeze_view_33", "pid": 0, "tid": 7,
    "ts": 1290508085037.261, "dur": 301.280,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 522.893616,      "warps per SM": 4183.148926,      "grid": [98304, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508085037.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508085341.133, "dur": 0.480,
    "args": {
      "External id": 223,      "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "bytes": 192, "memory bandwidth (GB/s)": 0.4
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508085341.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085342.637, "dur": 137.824,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 1.021277,      "warps per SM": 8.170213,      "grid": [64, 1, 3],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508085342.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508085482.541, "dur": 1.216,
    "args": {
      "External id": 223,      "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "bytes": 576, "memory bandwidth (GB/s)": 0.47368421052631576
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508085482.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085485.357, "dur": 140.448,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 222,      "shared memory": 73728,      "blocks per SM": 2.042553,      "warps per SM": 16.340425,      "grid": [192, 1, 2],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508085485.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_add_clamp_ne_sum_view_34", "pid": 0, "tid": 7,
    "ts": 1290508085626.125, "dur": 3.328,
    "args": {
      "External id": 223,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508756,      "registers per thread": 40,      "shared memory": 128,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508756, "pid": 0, "tid": 7, "ts": 1290508085626.125,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14565,
    "ts": 1290508014507.140, "dur": 366.450,
    "args": {
      "External id": 223,      "cbid": 311, "correlation": 508756
    }
  },
  {
    "ph": "s", "id": 508756, "pid": 14565, "tid": 14565, "ts": 1290508014507.140,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1290508085630.797, "dur": 1.536,
    "args": {
      "External id": 240,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508769,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508769, "pid": 0, "tid": 7, "ts": 1290508085630.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14565,
    "ts": 1290508015675.594, "dur": 13.117,
    "args": {
      "External id": 240,      "cbid": 211, "correlation": 508769
    }
  },
  {
    "ph": "s", "id": 508769, "pid": 14565, "tid": 14565, "ts": 1290508015675.594,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 14565, "tid": 14565,
    "ts": 1290508021481.131, "dur": 74642.637,
    "args": {
            "cbid": 165, "correlation": 509591
    }
  },
  {
    "ph": "s", "id": 509591, "pid": 14565, "tid": 14565, "ts": 1290508021481.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507954142.771, "dur": 7.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954142.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507954150.035, "dur": 4.832,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954150.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507954155.027, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954155.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507954156.659, "dur": 12.608,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954156.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507954169.395, "dur": 7.040,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954169.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507954176.563, "dur": 1.728,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954176.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507954178.387, "dur": 7.584,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954178.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507954186.067, "dur": 8.896,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954186.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507954195.091, "dur": 5.824,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954195.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954201.043, "dur": 22.592,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954201.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507954223.731, "dur": 3.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954223.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954226.963, "dur": 27.168,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954226.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507954254.259, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954254.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507954255.891, "dur": 25.344,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954255.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507954281.363, "dur": 6.816,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954281.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507954288.307, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954288.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954289.683, "dur": 23.520,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954289.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507954313.331, "dur": 3.200,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954313.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954316.659, "dur": 30.112,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954316.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507954346.899, "dur": 1.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954346.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507954348.627, "dur": 32.000,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954348.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507954383.955, "dur": 1.184,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.12162162162162163
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954383.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954390.003, "dur": 50.368,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954390.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954440.467, "dur": 48.064,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954440.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507954488.659, "dur": 2.080,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954488.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507954490.867, "dur": 21.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954490.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507954512.243, "dur": 6.656,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954512.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507954518.995, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954518.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954520.371, "dur": 13.280,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954520.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507954533.779, "dur": 2.848,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954533.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954536.723, "dur": 16.384,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954536.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507954553.235, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954553.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507954554.611, "dur": 8.928,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954554.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507954563.667, "dur": 3.520,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954563.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507954567.347, "dur": 1.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954567.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507954568.499, "dur": 14.912,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954568.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507954583.539, "dur": 1.120,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954583.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954584.787, "dur": 13.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954584.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507954598.419, "dur": 2.176,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954598.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954600.723, "dur": 3.616,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954600.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507954604.467, "dur": 0.960,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954604.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507954605.555, "dur": 159.776,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954605.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507954765.491, "dur": 7.200,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954765.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507954772.851, "dur": 4.896,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954772.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507954777.875, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954777.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507954779.507, "dur": 11.648,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954779.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507954791.283, "dur": 7.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954791.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507954798.547, "dur": 1.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954798.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507954800.339, "dur": 7.680,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954800.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507954808.147, "dur": 9.824,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954808.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507954818.067, "dur": 5.984,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954818.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954824.179, "dur": 22.752,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954824.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507954847.027, "dur": 3.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954847.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954850.259, "dur": 27.168,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954850.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507954877.555, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954877.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507954879.219, "dur": 24.352,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954879.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507954903.699, "dur": 6.880,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954903.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507954910.739, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954910.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954912.083, "dur": 22.784,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954912.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507954934.995, "dur": 3.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954934.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507954938.195, "dur": 30.144,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954938.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507954968.435, "dur": 1.632,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954968.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507954970.195, "dur": 31.552,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507954970.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507955004.371, "dur": 1.056,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955004.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955010.291, "dur": 51.456,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955010.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955061.875, "dur": 48.352,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955061.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507955110.323, "dur": 2.080,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955110.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507955112.531, "dur": 21.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955112.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507955134.355, "dur": 6.688,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955134.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955141.171, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955141.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955142.547, "dur": 13.792,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955142.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507955156.435, "dur": 2.880,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955156.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955159.443, "dur": 16.512,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955159.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507955176.083, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955176.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507955177.459, "dur": 9.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955177.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507955187.219, "dur": 3.232,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955187.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507955190.579, "dur": 1.088,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955190.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507955191.795, "dur": 15.328,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955191.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507955207.251, "dur": 1.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955207.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955208.435, "dur": 13.952,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955208.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507955222.547, "dur": 2.240,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955222.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955224.915, "dur": 3.712,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955224.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507955228.755, "dur": 0.960,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955228.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507955229.843, "dur": 160.256,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955229.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507955390.227, "dur": 7.392,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955390.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507955397.747, "dur": 4.800,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955397.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507955402.675, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955402.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507955404.339, "dur": 12.160,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955404.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507955416.627, "dur": 7.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955416.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507955423.891, "dur": 1.664,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955423.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507955425.715, "dur": 7.456,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955425.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507955433.331, "dur": 9.152,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955433.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507955442.611, "dur": 6.112,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955442.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955448.851, "dur": 23.168,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955448.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507955472.147, "dur": 3.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955472.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955475.411, "dur": 26.816,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955475.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507955502.323, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955502.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290507955503.987, "dur": 20.736,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955503.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507955524.819, "dur": 7.008,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955524.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955531.923, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955531.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290507955533.267, "dur": 18.528,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955533.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290507955551.891, "dur": 13.440,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955551.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955565.427, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955565.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955566.771, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955566.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290507955568.147, "dur": 25.376,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955568.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955593.651, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955593.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955594.963, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955594.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290507955596.307, "dur": 1.024,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955596.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290507955597.459, "dur": 36.992,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955597.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290507955634.579, "dur": 9.856,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955634.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955644.563, "dur": 1.888,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955644.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955646.579, "dur": 10.688,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955646.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507955657.395, "dur": 2.944,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955657.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955660.467, "dur": 11.936,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955660.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507955672.531, "dur": 1.152,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955672.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290507955673.811, "dur": 7.616,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955673.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955681.523, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955681.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290507955682.899, "dur": 4.480,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955682.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955687.475, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955687.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955688.819, "dur": 10.112,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955688.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507955699.059, "dur": 2.880,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955699.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955702.035, "dur": 10.432,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955702.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507955712.563, "dur": 1.152,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955712.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290507955713.843, "dur": 2.464,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955713.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290507955716.435, "dur": 7.424,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955716.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955723.955, "dur": 23.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955723.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507955747.187, "dur": 3.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955747.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955750.419, "dur": 29.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955750.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507955780.243, "dur": 1.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955780.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507955782.003, "dur": 32.576,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955782.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507955816.499, "dur": 1.024,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955816.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955822.387, "dur": 53.376,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955822.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955875.891, "dur": 47.712,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955875.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507955923.763, "dur": 2.080,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955923.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507955925.971, "dur": 29.952,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955925.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507955956.019, "dur": 6.848,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955956.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507955962.995, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955962.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955964.339, "dur": 13.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955964.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507955977.683, "dur": 2.816,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955977.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507955980.595, "dur": 15.936,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955980.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507955996.659, "dur": 1.280,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955996.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507955998.067, "dur": 9.472,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507955998.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507956007.667, "dur": 3.168,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956007.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507956010.963, "dur": 1.088,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956010.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507956012.179, "dur": 15.072,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956012.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507956027.379, "dur": 1.024,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956027.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956028.499, "dur": 13.440,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956028.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507956042.067, "dur": 2.240,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956042.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956044.403, "dur": 3.744,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956044.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507956048.275, "dur": 0.992,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956048.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507956049.395, "dur": 158.560,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956049.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507956208.147, "dur": 7.008,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956208.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507956215.251, "dur": 4.896,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956215.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507956220.275, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956220.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507956221.907, "dur": 12.704,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956221.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507956234.739, "dur": 6.816,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956234.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507956241.683, "dur": 1.792,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956241.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507956243.571, "dur": 7.552,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956243.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507956251.251, "dur": 9.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956251.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507956260.659, "dur": 6.016,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956260.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956266.803, "dur": 22.944,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956266.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507956289.875, "dur": 3.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956289.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956293.107, "dur": 26.912,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956293.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507956320.179, "dur": 1.472,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956320.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507956321.811, "dur": 23.584,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956321.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507956345.491, "dur": 6.784,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956345.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507956352.371, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956352.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956353.715, "dur": 22.976,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956353.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507956376.787, "dur": 3.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956376.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956380.019, "dur": 29.760,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956380.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507956409.907, "dur": 1.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956409.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507956411.731, "dur": 31.808,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956411.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507956445.171, "dur": 1.056,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956445.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956451.091, "dur": 49.952,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956451.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956501.171, "dur": 47.392,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956501.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507956548.659, "dur": 2.080,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956548.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507956550.867, "dur": 20.704,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956550.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507956571.699, "dur": 6.560,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956571.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507956578.387, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956578.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956579.731, "dur": 13.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956579.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507956592.915, "dur": 2.848,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956592.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956595.891, "dur": 16.576,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956595.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507956612.595, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956612.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507956613.971, "dur": 9.184,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956613.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507956623.283, "dur": 3.040,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956623.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507956626.451, "dur": 1.024,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956626.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507956627.635, "dur": 15.392,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956627.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507956643.155, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956643.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956644.787, "dur": 13.440,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956644.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507956658.355, "dur": 2.208,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956658.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956660.659, "dur": 3.616,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956660.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507956664.403, "dur": 0.960,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956664.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507956665.459, "dur": 153.632,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956665.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507956819.219, "dur": 7.072,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956819.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507956826.387, "dur": 4.832,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956826.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507956831.315, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956831.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507956832.979, "dur": 11.616,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956832.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507956844.723, "dur": 6.848,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956844.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507956851.667, "dur": 1.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956851.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507956853.491, "dur": 7.872,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956853.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507956861.491, "dur": 9.472,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956861.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507956871.123, "dur": 5.952,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956871.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956877.235, "dur": 23.232,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956877.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507956900.627, "dur": 3.040,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956900.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956903.795, "dur": 26.752,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956903.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507956930.675, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956930.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507956932.307, "dur": 25.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956932.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507956957.907, "dur": 7.008,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956957.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507956965.011, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956965.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956966.355, "dur": 23.968,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956966.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507956990.419, "dur": 3.200,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956990.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507956993.715, "dur": 29.888,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507956993.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507957023.731, "dur": 1.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957023.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507957025.427, "dur": 31.840,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957025.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507957059.667, "dur": 1.025,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.1404878048780488
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957059.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957065.491, "dur": 50.720,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957065.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957116.371, "dur": 48.576,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957116.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507957165.107, "dur": 2.080,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957165.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507957167.316, "dur": 22.463,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957167.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507957189.875, "dur": 6.784,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957189.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507957196.787, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957196.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957198.163, "dur": 13.344,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957198.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507957211.603, "dur": 2.912,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957211.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957214.611, "dur": 16.032,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957214.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507957230.771, "dur": 1.249,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957230.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507957232.147, "dur": 9.088,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957232.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507957241.363, "dur": 3.137,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957241.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507957244.627, "dur": 1.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957244.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507957245.811, "dur": 14.752,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957245.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507957260.691, "dur": 1.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957260.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957261.843, "dur": 13.728,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957261.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507957275.667, "dur": 2.272,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957275.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957278.067, "dur": 3.648,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957278.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507957281.843, "dur": 0.961,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957281.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507957282.899, "dur": 163.328,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957282.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507957446.355, "dur": 7.137,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957446.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507957453.619, "dur": 4.896,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957453.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507957458.644, "dur": 1.503,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957458.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507957460.275, "dur": 11.680,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957460.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507957472.115, "dur": 6.944,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957472.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507957479.187, "dur": 1.729,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957479.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507957481.043, "dur": 8.097,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957481.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507957489.267, "dur": 10.721,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957489.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507957500.115, "dur": 6.112,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957500.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957506.356, "dur": 23.455,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957506.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507957529.940, "dur": 3.167,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957529.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957533.236, "dur": 26.944,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957533.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507957560.307, "dur": 1.505,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957560.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507957561.939, "dur": 23.328,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957561.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507957585.363, "dur": 6.785,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957585.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507957592.243, "dur": 1.249,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957592.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957593.588, "dur": 23.359,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957593.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507957617.076, "dur": 3.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957617.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957620.339, "dur": 30.112,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957620.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507957650.548, "dur": 1.663,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957650.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507957652.307, "dur": 31.552,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957652.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507957686.323, "dur": 0.992,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957686.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957692.180, "dur": 51.775,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957692.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957744.084, "dur": 47.359,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957744.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507957791.572, "dur": 2.015,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957791.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507957793.716, "dur": 21.695,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957793.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507957815.540, "dur": 6.559,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957815.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507957822.228, "dur": 1.247,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957822.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957823.603, "dur": 13.089,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957823.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507957836.819, "dur": 2.912,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957836.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957839.828, "dur": 16.447,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957839.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507957856.371, "dur": 1.280,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957856.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507957857.747, "dur": 9.408,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957857.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507957867.284, "dur": 3.167,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957867.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507957870.580, "dur": 1.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957870.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507957871.764, "dur": 14.463,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957871.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507957886.356, "dur": 0.991,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957886.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957887.476, "dur": 13.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957887.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507957901.203, "dur": 2.209,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957901.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507957903.539, "dur": 3.585,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957903.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507957907.252, "dur": 0.959,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957907.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507957908.308, "dur": 150.623,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507957908.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507958059.091, "dur": 6.881,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958059.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507958066.099, "dur": 4.833,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958066.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507958071.092, "dur": 1.503,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958071.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507958072.691, "dur": 13.057,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958072.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507958085.844, "dur": 6.719,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958085.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507958092.692, "dur": 1.760,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958092.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507958094.579, "dur": 7.392,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958094.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507958102.100, "dur": 8.415,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958102.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507958110.612, "dur": 5.888,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958110.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958116.596, "dur": 22.208,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958116.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507958138.932, "dur": 3.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958138.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958142.196, "dur": 27.008,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958142.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507958169.331, "dur": 1.473,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958169.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507958170.964, "dur": 24.000,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958170.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507958195.124, "dur": 6.912,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958195.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507958202.163, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958202.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958203.539, "dur": 22.945,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958203.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507958226.612, "dur": 3.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958226.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958229.843, "dur": 29.856,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958229.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507958259.828, "dur": 1.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958259.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507958261.524, "dur": 31.616,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958261.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507958294.452, "dur": 1.024,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958294.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958300.948, "dur": 53.120,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958300.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958354.196, "dur": 48.416,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958354.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507958402.740, "dur": 2.048,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958402.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507958404.948, "dur": 21.472,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958404.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507958426.548, "dur": 6.655,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958426.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507958433.300, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958433.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958434.676, "dur": 13.312,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958434.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507958448.116, "dur": 2.816,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958448.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958451.060, "dur": 16.736,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958451.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507958467.891, "dur": 1.249,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958467.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507958469.268, "dur": 9.632,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958469.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507958479.028, "dur": 3.200,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958479.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507958482.388, "dur": 1.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958482.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507958483.572, "dur": 15.392,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958483.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507958499.092, "dur": 1.120,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958499.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958500.372, "dur": 13.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958500.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507958514.100, "dur": 2.239,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958514.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958516.468, "dur": 3.647,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958516.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507958520.244, "dur": 0.960,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958520.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507958521.300, "dur": 167.840,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958521.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507958689.268, "dur": 7.232,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958689.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507958696.628, "dur": 4.768,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958696.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507958701.492, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958701.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507958703.124, "dur": 11.648,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958703.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507958714.868, "dur": 6.720,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958714.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507958721.716, "dur": 1.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958721.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507958723.540, "dur": 8.800,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958723.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507958732.436, "dur": 10.976,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958732.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507958743.540, "dur": 5.952,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958743.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958749.620, "dur": 22.304,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958749.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507958772.084, "dur": 3.008,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958772.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958775.252, "dur": 26.720,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958775.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507958802.132, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958802.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507958803.764, "dur": 23.040,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958803.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507958826.900, "dur": 6.976,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958826.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507958834.004, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958834.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958835.348, "dur": 22.816,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958835.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507958858.260, "dur": 3.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958858.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958861.460, "dur": 29.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958861.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507958891.156, "dur": 1.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958891.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507958892.852, "dur": 31.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958892.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507958926.292, "dur": 1.152,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958926.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958932.276, "dur": 49.792,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958932.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507958982.196, "dur": 47.584,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507958982.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507959029.908, "dur": 2.048,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959029.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507959032.084, "dur": 21.184,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959032.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507959053.396, "dur": 6.656,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959053.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507959060.180, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959060.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959061.492, "dur": 13.184,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959061.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507959074.804, "dur": 2.880,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959074.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959077.780, "dur": 16.544,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959077.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507959094.420, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959094.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507959095.796, "dur": 9.312,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959095.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507959105.236, "dur": 3.200,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959105.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507959108.564, "dur": 1.088,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959108.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507959109.780, "dur": 15.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959109.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507959125.012, "dur": 1.024,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959125.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959126.164, "dur": 13.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959126.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507959139.988, "dur": 2.208,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959139.988,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959142.324, "dur": 3.712,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959142.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507959146.164, "dur": 0.928,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959146.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507959147.220, "dur": 150.464,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959147.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507959297.844, "dur": 7.072,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959297.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507959305.044, "dur": 4.832,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959305.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507959309.972, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959309.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507959311.604, "dur": 12.256,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959311.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507959323.956, "dur": 6.752,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959323.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507959330.836, "dur": 1.696,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959330.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507959332.660, "dur": 7.520,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959332.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507959340.308, "dur": 9.728,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959340.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507959350.196, "dur": 6.528,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959350.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959356.852, "dur": 22.624,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959356.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507959379.604, "dur": 3.104,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959379.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959382.836, "dur": 27.296,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959382.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507959410.260, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959410.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507959411.924, "dur": 24.288,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959411.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507959436.308, "dur": 6.720,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959436.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507959443.124, "dur": 1.280,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959443.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959444.500, "dur": 22.880,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959444.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507959467.476, "dur": 3.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959467.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959470.708, "dur": 29.920,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959470.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507959500.756, "dur": 1.568,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959500.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507959502.484, "dur": 32.448,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959502.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507959536.596, "dur": 0.992,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959536.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959542.420, "dur": 52.128,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959542.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959594.676, "dur": 48.352,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959594.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507959643.156, "dur": 2.048,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959643.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507959645.332, "dur": 20.896,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959645.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507959666.356, "dur": 6.688,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959666.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507959673.172, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959673.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959674.548, "dur": 13.632,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959674.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507959688.308, "dur": 2.880,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959688.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959691.316, "dur": 16.928,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959691.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507959708.372, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959708.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507959709.716, "dur": 9.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959709.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507959719.348, "dur": 3.168,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959719.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507959722.644, "dur": 1.088,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959722.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507959723.828, "dur": 14.656,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959723.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507959738.612, "dur": 1.728,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959738.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959740.468, "dur": 13.920,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959740.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507959754.516, "dur": 2.240,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959754.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959756.884, "dur": 3.552,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959756.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507959760.596, "dur": 0.928,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959760.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507959761.652, "dur": 155.968,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959761.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507959917.780, "dur": 6.976,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959917.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507959924.884, "dur": 5.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959924.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507959930.260, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959930.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507959931.924, "dur": 11.776,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959931.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507959943.828, "dur": 7.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959943.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507959951.060, "dur": 1.728,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959951.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507959952.884, "dur": 7.392,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959952.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507959960.372, "dur": 9.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959960.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507959969.716, "dur": 5.792,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959969.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507959975.636, "dur": 22.016,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959975.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507959997.780, "dur": 3.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507959997.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960001.012, "dur": 26.784,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960001.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507960027.892, "dur": 1.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960027.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290507960029.524, "dur": 23.136,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960029.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507960052.788, "dur": 6.976,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960052.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960059.892, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960059.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290507960061.236, "dur": 18.592,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960061.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290507960079.924, "dur": 13.408,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960079.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960093.460, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960093.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960094.804, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960094.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290507960096.180, "dur": 25.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960096.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960121.780, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960121.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960123.092, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960123.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290507960124.436, "dur": 1.024,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960124.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290507960125.588, "dur": 37.824,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960125.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290507960163.540, "dur": 13.280,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960163.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960176.948, "dur": 1.216,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960176.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960178.292, "dur": 10.688,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960178.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507960189.076, "dur": 2.912,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960189.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960192.116, "dur": 11.904,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960192.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507960204.148, "dur": 1.184,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960204.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290507960205.428, "dur": 7.552,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960205.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960213.076, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960213.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290507960214.452, "dur": 4.448,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960214.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960219.028, "dur": 1.184,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960219.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960220.372, "dur": 10.272,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960220.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507960230.772, "dur": 2.816,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960230.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960233.716, "dur": 10.432,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960233.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507960244.276, "dur": 1.152,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960244.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290507960245.556, "dur": 2.528,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960245.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290507960248.212, "dur": 7.872,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960248.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960256.212, "dur": 22.944,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960256.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507960279.284, "dur": 3.008,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960279.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960282.420, "dur": 29.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960282.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507960312.148, "dur": 1.600,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960312.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507960313.876, "dur": 33.152,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960313.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507960348.692, "dur": 0.992,
    "args": {
            "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960348.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960354.612, "dur": 55.296,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960354.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960410.004, "dur": 48.384,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960410.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507960458.548, "dur": 2.112,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960458.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290507960460.788, "dur": 33.760,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960460.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960494.676, "dur": 16.800,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960494.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507960511.604, "dur": 9.824,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960511.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507960521.556, "dur": 15.808,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960521.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960537.492, "dur": 5.280,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960537.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507960542.868, "dur": 164.704,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960542.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507960707.732, "dur": 13.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960707.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507960721.396, "dur": 7.424,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960721.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507960728.948, "dur": 6.144,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960728.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507960735.220, "dur": 4.928,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960735.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507960740.276, "dur": 5.568,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960740.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960745.972, "dur": 27.264,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960745.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290507960773.364, "dur": 19.936,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960773.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960793.428, "dur": 1.248,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960793.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960794.772, "dur": 13.856,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960794.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507960808.724, "dur": 2.848,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960808.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507960811.700, "dur": 1.856,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960811.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507960813.652, "dur": 3.328,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960813.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507960817.108, "dur": 1.056,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960817.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507960818.292, "dur": 1.472,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960818.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960819.892, "dur": 13.504,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960819.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507960833.556, "dur": 2.080,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960833.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507960835.796, "dur": 0.928,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960835.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507960836.852, "dur": 4.896,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960836.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507960841.908, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960841.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507960843.572, "dur": 7.392,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960843.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507960851.092, "dur": 1.728,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960851.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507960852.916, "dur": 22.464,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960852.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507960875.476, "dur": 2.976,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960875.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507960878.580, "dur": 1.536,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960878.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290507960880.244, "dur": 24.896,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960880.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507960905.268, "dur": 1.440,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 501725,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 501725, "pid": 0, "tid": 7, "ts": 1290507960905.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290507960910.004, "dur": 8.608,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502502,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502502, "pid": 0, "tid": 7, "ts": 1290507960910.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290507960918.964, "dur": 15.584,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502502,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502502, "pid": 0, "tid": 7, "ts": 1290507960918.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290507960936.052, "dur": 22.560,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502513,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502513, "pid": 0, "tid": 7, "ts": 1290507960936.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290507967381.238, "dur": 2.016,
    "args": {
      "External id": 516,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502598,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502598, "pid": 0, "tid": 7, "ts": 1290507967381.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507958550.313, "dur": 13.984,
    "args": {
      "External id": 516,      "cbid": 211, "correlation": 502598
    }
  },
  {
    "ph": "s", "id": 502598, "pid": 14565, "tid": 14640, "ts": 1290507958550.313,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507958572.203, "dur": 0.700,
    "args": {
      "External id": 514,      "cbid": 317, "correlation": 502606
    }
  },
  {
    "ph": "f", "id": 502606, "pid": 14565, "tid": 14640, "ts": 1290507958572.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290507967387.477, "dur": 12.480,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967387.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290507967400.086, "dur": 17.887,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967400.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290507967418.069, "dur": 6.496,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967418.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507967424.661, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967424.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967426.005, "dur": 23.393,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967426.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507967449.525, "dur": 3.232,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967449.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967452.853, "dur": 30.273,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967452.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507967483.253, "dur": 1.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967483.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507967484.982, "dur": 33.407,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967484.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507967520.790, "dur": 1.023,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.14076246334310852
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967520.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967526.709, "dur": 56.320,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967526.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967583.158, "dur": 48.383,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967583.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507967631.670, "dur": 2.079,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967631.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290507967633.846, "dur": 30.079,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967633.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507967664.053, "dur": 6.433,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967664.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507967670.614, "dur": 1.247,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967670.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967671.990, "dur": 13.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967671.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507967685.397, "dur": 2.912,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967685.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967688.406, "dur": 16.063,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967688.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507967704.598, "dur": 1.247,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967704.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507967706.005, "dur": 9.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967706.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507967715.382, "dur": 3.200,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967715.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507967718.709, "dur": 1.089,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967718.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507967719.894, "dur": 14.911,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967719.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507967734.901, "dur": 1.057,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967734.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967736.085, "dur": 13.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967736.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507967749.750, "dur": 2.240,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967749.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967752.117, "dur": 3.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967752.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507967755.893, "dur": 0.929,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967755.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507967756.949, "dur": 163.744,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967756.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507967920.822, "dur": 7.039,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967920.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507967927.990, "dur": 4.799,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967927.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507967932.917, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967932.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507967934.549, "dur": 11.584,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967934.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507967946.262, "dur": 7.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967946.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507967953.398, "dur": 1.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967953.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507967955.221, "dur": 7.617,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967955.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507967962.965, "dur": 8.481,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967962.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507967971.606, "dur": 5.888,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967971.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507967977.622, "dur": 22.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967977.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507967999.957, "dur": 3.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507967999.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968003.189, "dur": 26.913,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968003.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507968030.229, "dur": 1.473,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968030.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507968031.829, "dur": 19.905,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968031.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507968051.861, "dur": 6.721,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968051.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507968058.709, "dur": 1.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968058.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968060.085, "dur": 23.073,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968060.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507968083.253, "dur": 3.041,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968083.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968086.389, "dur": 30.112,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968086.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507968116.598, "dur": 1.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968116.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507968118.326, "dur": 32.447,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968118.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507968153.526, "dur": 0.992,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968153.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968159.350, "dur": 49.695,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968159.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968209.142, "dur": 48.384,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968209.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507968257.686, "dur": 2.047,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968257.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507968259.893, "dur": 21.793,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968259.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507968281.814, "dur": 6.688,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968281.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507968288.598, "dur": 1.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968288.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968289.974, "dur": 13.535,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968289.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507968303.606, "dur": 2.879,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968303.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968306.614, "dur": 16.576,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968306.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507968323.317, "dur": 1.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968323.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507968324.726, "dur": 9.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968324.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507968334.358, "dur": 2.912,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968334.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507968337.397, "dur": 1.089,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968337.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507968338.614, "dur": 15.296,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968338.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507968354.038, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968354.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968355.190, "dur": 13.856,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968355.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507968369.174, "dur": 2.240,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968369.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968371.509, "dur": 3.649,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968371.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507968375.318, "dur": 0.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968375.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507968376.374, "dur": 169.344,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968376.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507968545.846, "dur": 7.392,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968545.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507968553.366, "dur": 4.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968553.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507968558.293, "dur": 1.505,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968558.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507968559.926, "dur": 12.096,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968559.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507968572.150, "dur": 6.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968572.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507968579.254, "dur": 1.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968579.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507968581.046, "dur": 7.552,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968581.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507968588.694, "dur": 10.016,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968588.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507968598.838, "dur": 6.176,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968598.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968605.142, "dur": 22.656,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968605.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507968627.926, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968627.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968631.126, "dur": 26.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968631.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507968658.070, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968658.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507968659.734, "dur": 25.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968659.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507968685.142, "dur": 6.720,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968685.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507968691.990, "dur": 1.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968691.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968693.366, "dur": 21.984,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968693.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507968715.478, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968715.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968718.678, "dur": 29.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968718.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507968748.438, "dur": 1.664,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968748.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507968750.230, "dur": 31.904,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968750.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507968784.374, "dur": 1.056,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968784.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968790.294, "dur": 53.856,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968790.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968844.278, "dur": 47.392,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968844.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507968891.830, "dur": 2.112,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968891.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507968894.070, "dur": 21.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968894.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507968915.990, "dur": 6.592,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968915.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507968922.678, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968922.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968924.054, "dur": 13.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968924.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507968937.366, "dur": 2.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968937.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968940.310, "dur": 16.384,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968940.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507968956.790, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968956.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507968958.134, "dur": 9.568,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968958.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507968967.830, "dur": 2.784,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968967.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507968970.742, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968970.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507968971.926, "dur": 15.424,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968971.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507968987.478, "dur": 1.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968987.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507968989.206, "dur": 13.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507968989.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507969002.966, "dur": 2.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969002.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969005.302, "dur": 3.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969005.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507969009.046, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969009.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507969010.134, "dur": 159.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969010.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507969169.430, "dur": 7.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969169.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507969176.566, "dur": 4.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969176.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507969181.622, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969181.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507969183.254, "dur": 11.904,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969183.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507969195.286, "dur": 7.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969195.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507969202.486, "dur": 1.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969202.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507969204.310, "dur": 7.520,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969204.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507969211.958, "dur": 8.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969211.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507969220.886, "dur": 5.952,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969220.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969226.966, "dur": 23.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969226.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507969250.198, "dur": 3.200,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969250.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969253.494, "dur": 26.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969253.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507969280.502, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969280.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507969282.166, "dur": 25.152,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969282.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507969307.446, "dur": 6.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969307.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507969314.454, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969314.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969315.830, "dur": 23.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969315.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507969339.606, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969339.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969342.806, "dur": 29.952,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969342.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507969372.918, "dur": 1.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969372.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507969374.678, "dur": 32.000,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969374.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507969409.078, "dur": 1.088,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969409.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969414.966, "dur": 51.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969414.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969466.710, "dur": 48.320,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969466.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507969515.158, "dur": 2.048,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969515.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507969517.366, "dur": 21.440,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969517.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507969538.902, "dur": 6.656,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969538.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507969545.654, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969545.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969547.030, "dur": 13.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969547.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507969560.310, "dur": 2.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969560.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969563.286, "dur": 16.160,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969563.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507969579.542, "dur": 1.312,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969579.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507969580.982, "dur": 9.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969580.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507969590.326, "dur": 2.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969590.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507969593.430, "dur": 1.088,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969593.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507969594.614, "dur": 14.528,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969594.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507969609.238, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969609.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969610.390, "dur": 13.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969610.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507969624.310, "dur": 2.240,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969624.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969626.678, "dur": 3.776,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969626.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507969630.582, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969630.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507969631.670, "dur": 158.080,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969631.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507969789.942, "dur": 6.912,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969789.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507969797.014, "dur": 4.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969797.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507969801.942, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969801.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507969803.606, "dur": 11.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969803.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507969815.350, "dur": 7.168,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969815.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507969822.614, "dur": 1.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969822.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507969824.438, "dur": 7.712,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969824.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507969832.278, "dur": 9.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969832.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507969842.198, "dur": 5.952,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969842.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969848.246, "dur": 22.240,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969848.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507969870.614, "dur": 3.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969870.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969873.878, "dur": 26.944,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969873.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507969900.950, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969900.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507969902.614, "dur": 22.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969902.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507969925.686, "dur": 6.720,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969925.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507969932.502, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969932.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969933.846, "dur": 23.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969933.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507969956.950, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969956.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507969960.150, "dur": 29.728,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969960.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507969990.006, "dur": 1.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969990.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507969991.734, "dur": 31.680,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507969991.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507970025.494, "dur": 1.120,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970025.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970031.446, "dur": 51.744,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970031.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970083.318, "dur": 47.584,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970083.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507970131.030, "dur": 2.048,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970131.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507970133.206, "dur": 20.576,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970133.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507970153.910, "dur": 6.592,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970153.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507970160.630, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970160.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970161.974, "dur": 13.120,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970161.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507970175.190, "dur": 2.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970175.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970178.166, "dur": 16.608,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970178.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507970194.902, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970194.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507970196.246, "dur": 9.664,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970196.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507970206.038, "dur": 3.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970206.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507970209.174, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970209.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507970210.326, "dur": 15.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970210.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507970226.102, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970226.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970227.254, "dur": 13.440,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970227.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507970240.822, "dur": 2.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970240.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970243.158, "dur": 3.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970243.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507970246.902, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970246.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507970247.958, "dur": 158.112,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970247.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507970406.198, "dur": 7.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970406.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507970413.398, "dur": 4.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970413.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507970418.454, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970418.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507970420.118, "dur": 11.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970420.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507970431.894, "dur": 7.328,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970431.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507970439.350, "dur": 1.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970439.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507970441.174, "dur": 7.712,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970441.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507970449.014, "dur": 9.408,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970449.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507970458.518, "dur": 6.016,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970458.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970464.630, "dur": 22.816,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970464.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507970487.542, "dur": 3.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970487.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970490.774, "dur": 26.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970490.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507970517.782, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970517.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507970519.414, "dur": 28.480,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970519.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507970548.022, "dur": 6.816,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970548.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507970554.934, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970554.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970556.310, "dur": 23.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970556.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507970579.574, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970579.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970582.774, "dur": 29.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970582.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507970612.438, "dur": 1.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970612.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507970614.198, "dur": 31.424,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970614.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507970648.054, "dur": 1.120,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970648.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970654.070, "dur": 51.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970654.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970705.814, "dur": 48.448,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970705.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507970754.390, "dur": 2.048,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970754.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507970756.566, "dur": 20.832,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970756.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507970777.526, "dur": 6.592,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970777.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507970784.214, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970784.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970785.590, "dur": 13.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970785.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507970799.318, "dur": 2.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970799.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970802.326, "dur": 16.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970802.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507970819.254, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970819.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507970820.630, "dur": 9.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970820.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507970830.006, "dur": 3.040,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970830.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507970833.174, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970833.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507970834.358, "dur": 15.264,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970834.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507970849.750, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970849.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970850.902, "dur": 13.920,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970850.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507970864.918, "dur": 2.272,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970864.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507970867.286, "dur": 3.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970867.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507970871.062, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970871.062,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507970872.150, "dur": 156.096,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507970872.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507971028.406, "dur": 6.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971028.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507971035.542, "dur": 4.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971035.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507971040.502, "dur": 1.472,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971040.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507971042.134, "dur": 11.840,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971042.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507971054.102, "dur": 6.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971054.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507971061.206, "dur": 1.728,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971061.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507971063.030, "dur": 7.520,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971063.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507971070.646, "dur": 8.672,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971070.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507971079.414, "dur": 5.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971079.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971085.302, "dur": 22.496,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971085.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507971107.894, "dur": 3.168,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971107.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971111.158, "dur": 27.264,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971111.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507971138.550, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971138.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507971140.182, "dur": 25.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971140.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507971165.846, "dur": 6.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971165.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507971172.950, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971172.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971174.326, "dur": 23.488,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971174.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507971197.942, "dur": 3.040,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971197.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971201.110, "dur": 29.760,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971201.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507971230.998, "dur": 1.568,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971230.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507971232.726, "dur": 32.000,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971232.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507971266.582, "dur": 1.152,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971266.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971272.598, "dur": 54.176,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971272.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971326.870, "dur": 47.904,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971326.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507971374.870, "dur": 2.112,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971374.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507971377.078, "dur": 21.824,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971377.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507971398.998, "dur": 6.688,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971398.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507971405.814, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971405.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971407.158, "dur": 13.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971407.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507971420.566, "dur": 2.912,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971420.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971423.574, "dur": 16.416,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971423.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507971440.086, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971440.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507971441.462, "dur": 9.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971441.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507971450.870, "dur": 3.200,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971450.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507971454.230, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971454.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507971455.382, "dur": 14.656,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971455.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507971470.166, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971470.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971471.542, "dur": 13.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971471.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507971485.206, "dur": 2.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971485.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971487.542, "dur": 3.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971487.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507971491.254, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971491.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507971492.342, "dur": 160.416,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971492.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507971652.950, "dur": 7.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971652.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507971660.214, "dur": 4.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971660.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507971665.270, "dur": 1.472,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971665.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507971666.870, "dur": 11.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971666.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507971678.614, "dur": 7.392,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971678.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507971686.134, "dur": 1.728,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971686.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507971687.958, "dur": 7.712,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971687.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507971695.798, "dur": 10.016,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971695.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507971705.910, "dur": 6.528,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971705.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971712.566, "dur": 22.816,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971712.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507971735.510, "dur": 3.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971735.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971738.774, "dur": 27.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971738.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507971765.974, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971765.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507971767.606, "dur": 24.416,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971767.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507971792.182, "dur": 6.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971792.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507971799.190, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971799.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971800.534, "dur": 22.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971800.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507971823.510, "dur": 3.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971823.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971826.742, "dur": 29.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971826.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507971856.662, "dur": 1.568,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971856.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507971858.390, "dur": 31.584,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971858.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507971892.310, "dur": 1.088,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971892.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971898.230, "dur": 51.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971898.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507971949.430, "dur": 48.224,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971949.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507971997.782, "dur": 2.080,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507971997.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507972000.022, "dur": 21.408,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972000.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507972021.558, "dur": 6.624,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972021.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972028.310, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972028.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972029.654, "dur": 13.312,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972029.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972043.094, "dur": 2.784,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972043.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972046.006, "dur": 16.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972046.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507972063.062, "dur": 1.217,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972063.062,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507972064.406, "dur": 9.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972064.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507972074.230, "dur": 2.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972074.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507972077.334, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972077.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507972078.519, "dur": 15.199,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972078.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507972093.815, "dur": 1.151,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972093.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972095.094, "dur": 13.728,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972095.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972108.950, "dur": 2.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972108.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972111.286, "dur": 3.584,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972111.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507972114.998, "dur": 0.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972114.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507972116.054, "dur": 153.440,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972116.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507972269.622, "dur": 7.328,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972269.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507972277.078, "dur": 4.736,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972277.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507972281.942, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972281.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507972283.575, "dur": 11.647,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972283.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507972295.350, "dur": 7.264,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972295.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507972302.742, "dur": 1.665,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972302.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507972304.534, "dur": 7.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972304.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507972312.278, "dur": 9.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972312.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507972321.463, "dur": 6.048,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972321.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972327.638, "dur": 22.496,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972327.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972350.230, "dur": 3.168,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972350.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972353.526, "dur": 26.720,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972353.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507972380.342, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972380.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290507972382.006, "dur": 20.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972382.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507972403.094, "dur": 7.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972403.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972410.198, "dur": 1.249,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972410.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290507972411.542, "dur": 18.432,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972411.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290507972430.070, "dur": 13.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972430.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972443.703, "dur": 1.247,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972443.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972445.079, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972445.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290507972446.422, "dur": 25.952,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972446.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972472.502, "dur": 1.217,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972472.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972473.846, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972473.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290507972475.158, "dur": 1.057,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972475.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290507972476.342, "dur": 37.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972476.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290507972513.655, "dur": 9.951,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972513.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972523.734, "dur": 1.889,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972523.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972525.719, "dur": 11.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972525.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972536.854, "dur": 2.945,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972536.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972539.926, "dur": 12.384,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972539.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507972552.439, "dur": 1.151,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972552.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290507972553.719, "dur": 7.552,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972553.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972561.398, "dur": 1.249,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972561.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290507972562.743, "dur": 4.383,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972562.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972567.254, "dur": 1.185,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972567.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972568.566, "dur": 10.144,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972568.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972578.839, "dur": 2.911,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972578.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972581.846, "dur": 10.464,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972581.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507972592.406, "dur": 1.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972592.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290507972593.686, "dur": 2.529,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972593.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290507972596.310, "dur": 7.424,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972596.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972603.830, "dur": 22.785,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972603.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972626.742, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972626.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972629.942, "dur": 29.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972629.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507972659.767, "dur": 1.567,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972659.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507972661.463, "dur": 32.415,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972661.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507972697.111, "dur": 0.992,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972697.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972702.966, "dur": 54.529,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972702.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972757.655, "dur": 48.448,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972757.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507972806.231, "dur": 2.079,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972806.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507972808.438, "dur": 29.472,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972808.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507972838.039, "dur": 6.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972838.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507972844.982, "dur": 1.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972844.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972846.358, "dur": 13.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972846.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972859.991, "dur": 2.879,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972859.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972862.967, "dur": 16.192,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972862.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507972879.286, "dur": 1.249,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972879.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507972880.663, "dur": 9.215,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972880.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507972890.007, "dur": 3.263,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972890.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507972893.430, "dur": 1.089,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972893.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507972894.615, "dur": 15.680,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972894.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507972910.422, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972910.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972911.575, "dur": 13.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972911.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507972925.495, "dur": 2.272,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972925.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507972927.862, "dur": 3.681,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972927.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507972931.670, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972931.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507972932.759, "dur": 158.304,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507972932.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507973091.191, "dur": 6.975,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973091.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507973098.295, "dur": 4.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973098.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507973103.222, "dur": 1.505,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973103.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507973104.823, "dur": 12.639,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973104.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507973117.558, "dur": 6.752,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973117.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507973124.439, "dur": 1.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973124.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507973126.326, "dur": 7.521,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973126.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507973133.974, "dur": 8.961,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973133.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507973143.095, "dur": 5.824,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973143.095,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973149.047, "dur": 22.304,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973149.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507973171.478, "dur": 3.009,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973171.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973174.614, "dur": 26.624,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973174.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507973201.367, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973201.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507973203.031, "dur": 23.744,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973203.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507973226.903, "dur": 7.232,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973226.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507973234.231, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973234.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973235.606, "dur": 23.553,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973235.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507973259.255, "dur": 3.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973259.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973262.486, "dur": 29.665,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973262.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507973292.311, "dur": 1.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973292.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507973294.071, "dur": 32.064,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973294.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507973327.959, "dur": 0.992,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973327.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973333.782, "dur": 52.449,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973333.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973386.359, "dur": 47.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973386.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507973434.134, "dur": 2.081,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973434.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507973436.343, "dur": 20.768,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973436.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507973457.238, "dur": 6.689,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973457.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507973464.023, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973464.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973465.366, "dur": 13.505,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973465.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507973478.966, "dur": 2.816,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973478.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973481.911, "dur": 16.480,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973481.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507973498.518, "dur": 1.249,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973498.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507973499.895, "dur": 9.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973499.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507973509.559, "dur": 3.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973509.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507973512.823, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973512.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507973514.007, "dur": 15.264,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973514.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507973529.399, "dur": 1.408,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973529.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973530.935, "dur": 13.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973530.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507973544.663, "dur": 2.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973544.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973546.967, "dur": 3.712,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973546.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507973550.775, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973550.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507973551.831, "dur": 153.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973551.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507973705.143, "dur": 7.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973705.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507973712.247, "dur": 4.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973712.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507973717.303, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973717.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507973718.967, "dur": 11.712,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973718.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507973730.775, "dur": 6.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973730.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507973737.783, "dur": 1.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973737.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507973739.607, "dur": 8.000,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973739.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507973747.767, "dur": 9.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973747.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507973757.111, "dur": 6.240,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973757.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973763.511, "dur": 23.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973763.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507973786.679, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973786.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973789.847, "dur": 27.200,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973789.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507973817.175, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973817.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507973818.807, "dur": 25.408,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973818.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507973844.343, "dur": 7.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973844.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507973851.543, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973851.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973852.919, "dur": 23.360,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973852.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507973876.375, "dur": 3.168,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973876.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973879.639, "dur": 29.760,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973879.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507973909.495, "dur": 1.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973909.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507973911.255, "dur": 30.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973911.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507973945.431, "dur": 1.056,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973945.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507973951.255, "dur": 52.384,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507973951.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974003.767, "dur": 48.480,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974003.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507974052.407, "dur": 2.112,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974052.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507974054.647, "dur": 22.016,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974054.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507974076.759, "dur": 7.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974076.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507974083.959, "dur": 1.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974083.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974085.335, "dur": 13.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974085.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507974098.967, "dur": 2.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974098.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974101.975, "dur": 16.064,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974101.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507974118.167, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974118.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507974119.543, "dur": 9.440,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974119.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507974129.143, "dur": 3.328,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974129.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507974132.599, "dur": 1.088,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974132.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507974133.815, "dur": 15.360,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974133.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507974149.303, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974149.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974150.455, "dur": 13.824,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974150.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507974164.407, "dur": 2.240,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974164.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974166.775, "dur": 3.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974166.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507974170.551, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974170.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507974171.607, "dur": 161.920,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974171.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507974333.655, "dur": 7.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974333.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507974340.855, "dur": 4.832,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974340.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507974345.815, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974345.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507974347.447, "dur": 11.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974347.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507974359.255, "dur": 6.912,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974359.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507974366.295, "dur": 1.664,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974366.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507974368.087, "dur": 7.968,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974368.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507974376.183, "dur": 10.336,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974376.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507974386.647, "dur": 5.920,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974386.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974392.727, "dur": 22.560,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974392.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507974415.447, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974415.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974418.647, "dur": 27.264,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974418.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507974446.039, "dur": 1.568,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974446.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507974447.703, "dur": 23.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974447.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507974470.967, "dur": 6.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974470.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507974477.911, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974477.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974479.255, "dur": 22.080,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974479.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507974501.495, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974501.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974504.695, "dur": 29.568,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974504.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507974534.359, "dur": 1.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974534.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507974536.087, "dur": 31.808,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974536.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507974569.911, "dur": 1.024,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974569.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974575.863, "dur": 51.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974575.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974626.999, "dur": 48.256,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974626.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507974675.383, "dur": 2.048,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974675.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507974677.559, "dur": 21.312,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974677.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507974698.999, "dur": 6.560,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974698.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507974705.687, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974705.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974707.031, "dur": 13.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974707.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507974720.343, "dur": 2.912,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974720.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974723.383, "dur": 16.672,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974723.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507974740.183, "dur": 1.280,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974740.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507974741.591, "dur": 8.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974741.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507974750.679, "dur": 3.424,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974750.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507974754.231, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974754.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507974755.415, "dur": 15.040,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974755.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507974770.583, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974770.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974771.735, "dur": 13.920,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974771.735,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507974785.783, "dur": 2.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974785.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507974788.119, "dur": 3.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974788.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507974791.895, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974791.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507974792.983, "dur": 156.704,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974792.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507974949.783, "dur": 6.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974949.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507974956.887, "dur": 4.864,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974956.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507974961.911, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974961.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507974963.543, "dur": 12.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974963.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507974976.471, "dur": 6.816,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974976.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507974983.383, "dur": 1.888,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974983.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507974985.367, "dur": 7.520,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974985.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507974993.015, "dur": 8.480,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507974993.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507975001.623, "dur": 5.920,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975001.623,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975007.671, "dur": 22.720,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975007.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507975030.487, "dur": 3.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975030.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975033.751, "dur": 26.944,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975033.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507975060.855, "dur": 1.472,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975060.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507975062.487, "dur": 23.808,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975062.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507975086.423, "dur": 6.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975086.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507975093.399, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975093.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975094.775, "dur": 22.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975094.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507975117.783, "dur": 3.168,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975117.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975121.079, "dur": 29.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975121.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507975150.967, "dur": 1.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975150.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507975152.727, "dur": 32.192,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975152.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507975186.519, "dur": 0.992,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975186.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975192.343, "dur": 50.400,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975192.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975242.871, "dur": 48.288,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975242.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507975291.255, "dur": 2.144,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975291.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507975293.527, "dur": 21.312,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975293.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507975314.967, "dur": 6.624,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975314.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507975321.687, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975321.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975323.063, "dur": 13.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975323.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507975336.343, "dur": 2.880,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975336.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975339.319, "dur": 16.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975339.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507975356.375, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975356.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507975357.751, "dur": 10.016,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975357.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507975367.895, "dur": 3.200,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975367.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507975371.223, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975371.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507975372.407, "dur": 15.584,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975372.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507975388.119, "dur": 1.024,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975388.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975389.303, "dur": 13.632,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975389.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507975403.063, "dur": 2.272,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975403.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975405.463, "dur": 3.648,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975405.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507975409.239, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975409.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507975410.327, "dur": 160.320,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975410.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507975570.775, "dur": 7.200,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975570.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507975578.071, "dur": 4.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975578.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507975582.967, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975582.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507975584.599, "dur": 11.552,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975584.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507975596.247, "dur": 6.816,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975596.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507975603.191, "dur": 1.696,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975603.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507975604.983, "dur": 8.288,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975604.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507975613.399, "dur": 10.688,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975613.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507975624.247, "dur": 6.080,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975624.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975630.455, "dur": 23.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975630.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507975653.719, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975653.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975656.951, "dur": 26.688,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975656.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507975683.799, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975683.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507975685.431, "dur": 22.336,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975685.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507975707.895, "dur": 7.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975707.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507975714.999, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975714.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975716.375, "dur": 23.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975716.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507975739.511, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975739.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975742.711, "dur": 29.728,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975742.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507975772.535, "dur": 1.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975772.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507975774.263, "dur": 31.520,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975774.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507975808.119, "dur": 0.992,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975808.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975813.975, "dur": 51.552,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975813.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975865.655, "dur": 48.448,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975865.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507975914.231, "dur": 2.016,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975914.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507975916.407, "dur": 21.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975916.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507975937.751, "dur": 6.688,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975937.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507975944.567, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975944.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975945.911, "dur": 13.472,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975945.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507975959.511, "dur": 2.912,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975959.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507975962.519, "dur": 16.704,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975962.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507975979.351, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975979.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507975980.727, "dur": 9.120,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975980.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507975989.975, "dur": 2.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975989.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507975993.079, "dur": 1.088,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975993.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507975994.295, "dur": 15.744,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507975994.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507976010.167, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976010.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976011.319, "dur": 13.792,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976011.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507976025.239, "dur": 2.240,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976025.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976027.607, "dur": 3.616,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976027.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507976031.351, "dur": 0.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976031.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507976032.439, "dur": 159.712,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976032.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507976192.311, "dur": 7.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976192.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507976199.543, "dur": 4.800,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976199.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507976204.471, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976204.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507976206.103, "dur": 12.096,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976206.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507976218.327, "dur": 6.720,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976218.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507976225.207, "dur": 1.664,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976225.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507976227.031, "dur": 7.520,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976227.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507976234.679, "dur": 9.728,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976234.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507976244.535, "dur": 6.016,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976244.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976250.679, "dur": 23.200,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976250.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507976274.007, "dur": 3.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976274.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976277.271, "dur": 27.232,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976277.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507976304.631, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976304.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507976306.263, "dur": 24.352,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976306.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507976330.743, "dur": 6.752,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976330.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507976337.591, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976337.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976338.967, "dur": 23.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976338.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507976362.103, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976362.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976365.271, "dur": 29.920,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976365.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507976395.319, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976395.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507976397.015, "dur": 31.680,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976397.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507976430.519, "dur": 1.024,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976430.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976436.631, "dur": 51.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976436.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976487.831, "dur": 48.256,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976487.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507976536.215, "dur": 2.080,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976536.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507976538.423, "dur": 20.992,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976538.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507976559.543, "dur": 6.688,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976559.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507976566.359, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976566.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976567.735, "dur": 13.376,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976567.735,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507976581.239, "dur": 2.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976581.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976584.215, "dur": 16.640,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976584.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507976600.983, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976600.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507976602.359, "dur": 9.120,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976602.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507976611.607, "dur": 2.976,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976611.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507976614.711, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976614.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507976615.895, "dur": 15.296,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976615.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507976631.319, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976631.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976632.951, "dur": 13.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976632.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507976646.583, "dur": 2.208,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976646.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976648.919, "dur": 3.584,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976648.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507976652.631, "dur": 0.928,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976652.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507976653.719, "dur": 157.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976653.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507976811.223, "dur": 6.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976811.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507976818.199, "dur": 4.992,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976818.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507976823.319, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976823.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507976824.951, "dur": 11.744,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976824.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507976836.791, "dur": 7.008,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976836.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507976843.895, "dur": 1.984,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976843.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507976846.007, "dur": 7.744,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976846.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507976853.847, "dur": 9.120,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976853.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507976863.095, "dur": 6.176,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976863.095,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976869.399, "dur": 23.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976869.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507976892.599, "dur": 3.104,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976892.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507976895.799, "dur": 27.136,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976895.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507976923.063, "dur": 1.504,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976923.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290507976924.695, "dur": 22.784,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976924.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507976947.607, "dur": 6.944,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976947.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507976954.679, "dur": 1.217,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976954.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290507976956.023, "dur": 19.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976956.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290507976975.223, "dur": 13.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976975.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507976988.888, "dur": 1.247,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976988.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507976990.263, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976990.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290507976991.639, "dur": 25.952,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507976991.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507977017.687, "dur": 1.217,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977017.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507977019.031, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977019.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290507977020.375, "dur": 1.025,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977020.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290507977021.560, "dur": 37.695,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977021.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290507977059.384, "dur": 13.183,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977059.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507977072.344, "dur": 1.247,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977072.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977073.719, "dur": 10.784,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977073.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507977084.599, "dur": 2.944,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977084.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977087.639, "dur": 11.872,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977087.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507977099.607, "dur": 1.184,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977099.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290507977100.887, "dur": 7.585,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977100.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507977108.567, "dur": 1.248,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977108.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290507977109.943, "dur": 4.576,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977109.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507977114.647, "dur": 1.216,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977114.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977115.991, "dur": 10.368,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977115.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507977126.487, "dur": 2.848,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977126.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977129.495, "dur": 10.432,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977129.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507977140.055, "dur": 1.152,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977140.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290507977141.335, "dur": 2.592,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977141.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290507977144.023, "dur": 7.840,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977144.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977151.991, "dur": 23.264,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977151.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507977175.383, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977175.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977178.551, "dur": 29.473,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977178.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507977208.151, "dur": 1.600,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977208.151,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507977209.847, "dur": 32.544,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977209.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507977244.599, "dur": 0.992,
    "args": {
      "External id": 514,      "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977244.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977251.447, "dur": 53.857,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977251.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977305.431, "dur": 48.128,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977305.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507977353.687, "dur": 2.081,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977353.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290507977355.895, "dur": 32.896,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977355.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977388.920, "dur": 17.023,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977388.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507977406.072, "dur": 9.439,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977406.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507977415.640, "dur": 16.159,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977415.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977431.896, "dur": 5.119,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977431.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507977437.111, "dur": 153.089,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977437.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507977590.360, "dur": 13.440,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977590.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507977603.927, "dur": 7.360,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977603.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507977611.415, "dur": 6.112,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977611.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507977617.655, "dur": 4.864,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977617.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507977622.647, "dur": 5.697,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977622.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977628.471, "dur": 27.233,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977628.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290507977655.831, "dur": 20.032,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977655.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507977675.960, "dur": 1.279,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977675.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977677.336, "dur": 13.760,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977677.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507977691.223, "dur": 2.816,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977691.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507977694.168, "dur": 2.207,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977694.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507977696.504, "dur": 3.647,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977696.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507977700.280, "dur": 1.056,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977700.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507977701.463, "dur": 1.440,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977701.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977703.031, "dur": 13.440,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977703.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507977716.631, "dur": 2.081,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977716.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507977718.839, "dur": 0.960,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977718.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507977719.928, "dur": 4.831,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977719.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507977724.919, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977724.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507977726.584, "dur": 7.455,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977726.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507977734.168, "dur": 1.695,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977734.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507977735.959, "dur": 22.304,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977735.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507977758.360, "dur": 3.072,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977758.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507977761.559, "dur": 1.536,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977761.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290507977763.224, "dur": 24.768,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977763.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507977788.088, "dur": 1.471,
    "args": {
      "External id": 514,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 502608,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 502608, "pid": 0, "tid": 7, "ts": 1290507977788.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507958573.472, "dur": 977.307,
    "args": {
      "External id": 514,      "cbid": 311, "correlation": 502608
    }
  },
  {
    "ph": "s", "id": 502608, "pid": 14565, "tid": 14640, "ts": 1290507958573.472,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507962125.216, "dur": 1.332,
    "args": {
      "External id": 1282,      "cbid": 317, "correlation": 503383
    }
  },
  {
    "ph": "f", "id": 503383, "pid": 14565, "tid": 14640, "ts": 1290507962125.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290507977791.799, "dur": 8.513,
    "args": {
      "External id": 1282,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503385,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503385, "pid": 0, "tid": 7, "ts": 1290507977791.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290507977800.663, "dur": 16.032,
    "args": {
      "External id": 1282,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503385,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503385, "pid": 0, "tid": 7, "ts": 1290507977800.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507962127.425, "dur": 16.216,
    "args": {
      "External id": 1282,      "cbid": 311, "correlation": 503385
    }
  },
  {
    "ph": "s", "id": 503385, "pid": 14565, "tid": 14640, "ts": 1290507962127.425,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290507977817.848, "dur": 22.464,
    "args": {
      "External id": 1284,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503396,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503396, "pid": 0, "tid": 7, "ts": 1290507977817.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507962243.202, "dur": 12.979,
    "args": {
      "External id": 1284,      "cbid": 211, "correlation": 503396
    }
  },
  {
    "ph": "s", "id": 503396, "pid": 14565, "tid": 14640, "ts": 1290507962243.202,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290507984273.113, "dur": 2.016,
    "args": {
      "External id": 1292,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503481,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503481, "pid": 0, "tid": 7, "ts": 1290507984273.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507965887.303, "dur": 14.687,
    "args": {
      "External id": 1292,      "cbid": 211, "correlation": 503481
    }
  },
  {
    "ph": "s", "id": 503481, "pid": 14565, "tid": 14640, "ts": 1290507965887.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507965910.566, "dur": 0.617,
    "args": {
      "External id": 1290,      "cbid": 317, "correlation": 503489
    }
  },
  {
    "ph": "f", "id": 503489, "pid": 14565, "tid": 14640, "ts": 1290507965910.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290507984279.353, "dur": 12.544,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984279.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290507984292.025, "dur": 18.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984292.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290507984310.297, "dur": 6.560,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984310.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507984316.953, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984316.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984318.297, "dur": 23.872,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984318.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507984342.297, "dur": 3.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984342.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984345.593, "dur": 30.400,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984345.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507984376.121, "dur": 1.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984376.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507984377.881, "dur": 33.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984377.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507984412.729, "dur": 1.024,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984412.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984418.649, "dur": 56.032,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984418.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984474.809, "dur": 47.904,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984474.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507984522.809, "dur": 2.048,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984522.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290507984524.985, "dur": 29.824,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984524.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507984554.937, "dur": 6.432,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984554.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507984561.497, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984561.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984562.841, "dur": 13.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984562.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507984576.249, "dur": 2.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984576.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984579.161, "dur": 16.096,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984579.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507984595.417, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984595.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507984596.761, "dur": 9.408,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984596.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507984606.329, "dur": 3.360,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984606.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507984609.785, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984609.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507984610.969, "dur": 15.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984610.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507984626.233, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984626.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984627.385, "dur": 13.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984627.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507984641.113, "dur": 2.208,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984641.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984643.449, "dur": 3.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984643.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507984647.257, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984647.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507984648.409, "dur": 164.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984648.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507984813.529, "dur": 7.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984813.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507984820.665, "dur": 4.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984820.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507984825.721, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984825.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507984827.385, "dur": 12.160,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984827.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507984839.673, "dur": 7.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984839.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507984846.873, "dur": 1.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984846.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507984848.729, "dur": 7.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984848.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507984856.505, "dur": 8.992,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984856.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507984865.657, "dur": 5.920,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984865.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984871.705, "dur": 22.816,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984871.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507984894.617, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984894.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984897.849, "dur": 27.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984897.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507984924.985, "dur": 1.472,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984924.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507984926.617, "dur": 20.064,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984926.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507984946.841, "dur": 6.752,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984946.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507984953.721, "dur": 1.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984953.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984955.097, "dur": 23.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984955.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507984978.201, "dur": 3.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984978.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507984981.369, "dur": 29.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507984981.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507985011.193, "dur": 1.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985011.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507985012.953, "dur": 32.384,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985012.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507985047.481, "dur": 1.120,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985047.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985053.369, "dur": 51.808,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985053.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985105.273, "dur": 48.448,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985105.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507985153.881, "dur": 2.048,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985153.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507985156.089, "dur": 21.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985156.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507985177.785, "dur": 6.624,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985177.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507985184.505, "dur": 1.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985184.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985185.881, "dur": 13.408,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985185.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507985199.385, "dur": 2.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985199.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985202.393, "dur": 16.608,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985202.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507985219.161, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985219.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507985220.537, "dur": 9.888,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985220.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507985230.553, "dur": 2.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985230.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507985233.561, "dur": 1.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985233.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507985234.745, "dur": 14.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985234.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507985249.657, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985249.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985250.777, "dur": 13.792,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985250.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507985264.665, "dur": 2.304,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985264.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985267.065, "dur": 3.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985267.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507985270.873, "dur": 0.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985270.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507985271.961, "dur": 159.712,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985271.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507985431.801, "dur": 7.424,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985431.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507985439.353, "dur": 4.864,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985439.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507985444.345, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985444.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507985445.977, "dur": 12.192,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985445.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507985458.297, "dur": 7.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985458.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507985465.497, "dur": 1.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985465.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507985467.289, "dur": 7.488,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985467.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507985474.905, "dur": 9.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985474.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507985484.665, "dur": 5.952,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985484.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985490.745, "dur": 22.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985490.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507985513.017, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985513.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985516.249, "dur": 26.752,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985516.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507985543.129, "dur": 1.472,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985543.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507985544.729, "dur": 25.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985544.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507985569.913, "dur": 6.816,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985569.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507985576.857, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985576.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985578.201, "dur": 23.232,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985578.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507985601.561, "dur": 3.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985601.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985604.697, "dur": 29.952,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985604.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507985634.777, "dur": 1.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985634.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507985636.505, "dur": 31.584,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985636.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507985670.201, "dur": 1.056,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985670.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985676.153, "dur": 52.000,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985676.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985728.281, "dur": 47.776,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985728.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507985776.217, "dur": 2.112,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985776.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507985778.457, "dur": 21.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985778.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507985800.313, "dur": 6.656,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985800.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507985807.097, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985807.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985808.473, "dur": 13.376,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985808.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507985821.977, "dur": 2.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985821.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985824.921, "dur": 16.576,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985824.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507985841.593, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985841.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507985842.969, "dur": 9.856,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985842.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507985852.953, "dur": 3.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985852.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507985856.249, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985856.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507985857.433, "dur": 15.488,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985857.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507985873.017, "dur": 1.440,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985873.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985874.585, "dur": 13.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985874.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507985888.217, "dur": 2.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985888.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507985890.553, "dur": 3.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985890.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507985894.329, "dur": 0.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985894.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507985895.417, "dur": 156.288,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507985895.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507986051.865, "dur": 7.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986051.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507986059.129, "dur": 4.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986059.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507986064.185, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986064.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507986065.849, "dur": 12.032,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986065.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507986077.977, "dur": 7.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986077.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507986085.081, "dur": 1.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986085.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507986086.873, "dur": 7.392,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986086.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507986094.393, "dur": 8.576,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986094.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507986103.097, "dur": 6.112,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986103.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986109.337, "dur": 22.976,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986109.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507986132.409, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986132.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986135.609, "dur": 26.912,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986135.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507986162.649, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986162.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507986164.281, "dur": 25.120,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986164.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507986189.529, "dur": 6.880,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986189.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507986196.537, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986196.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986197.881, "dur": 23.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986197.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507986221.017, "dur": 3.200,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986221.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986224.345, "dur": 30.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986224.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507986254.617, "dur": 1.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986254.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507986256.345, "dur": 32.256,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986256.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507986290.681, "dur": 1.056,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986290.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986296.505, "dur": 52.192,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986296.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986348.825, "dur": 47.776,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986348.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507986396.729, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986396.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507986398.969, "dur": 21.120,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986398.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507986420.185, "dur": 6.656,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986420.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507986426.937, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986426.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986428.313, "dur": 13.152,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986428.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507986441.593, "dur": 2.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986441.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986444.569, "dur": 15.840,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986444.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507986460.537, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986460.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507986461.913, "dur": 9.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986461.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507986471.225, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986471.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507986474.457, "dur": 1.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986474.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507986475.641, "dur": 14.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986475.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507986490.681, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986490.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986491.833, "dur": 13.312,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986491.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507986505.273, "dur": 2.208,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986505.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986507.609, "dur": 3.616,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986507.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507986511.353, "dur": 0.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986511.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507986512.409, "dur": 160.160,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986512.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507986672.729, "dur": 6.976,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986672.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507986679.833, "dur": 4.800,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986679.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507986684.729, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986684.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507986686.361, "dur": 11.712,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986686.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507986698.169, "dur": 7.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986698.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507986705.369, "dur": 1.696,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986705.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507986707.193, "dur": 7.808,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986707.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507986715.129, "dur": 9.760,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986715.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507986724.985, "dur": 6.112,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986724.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986731.193, "dur": 22.304,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986731.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507986753.625, "dur": 3.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986753.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986756.825, "dur": 26.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986756.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507986783.801, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986783.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507986785.465, "dur": 23.936,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986785.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507986809.529, "dur": 6.752,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986809.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507986816.409, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986816.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986817.785, "dur": 23.328,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986817.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507986841.241, "dur": 3.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986841.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986844.441, "dur": 29.888,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986844.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507986874.457, "dur": 1.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986874.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507986876.185, "dur": 31.680,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986876.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507986909.241, "dur": 0.992,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986909.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986915.033, "dur": 50.752,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986915.033,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507986965.913, "dur": 48.736,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507986965.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507987014.777, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987014.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507987016.985, "dur": 21.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987016.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507987038.201, "dur": 6.688,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987038.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507987045.017, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987045.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987046.393, "dur": 13.185,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987046.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507987059.705, "dur": 2.880,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987059.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987062.714, "dur": 17.119,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987062.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507987079.961, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987079.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507987081.306, "dur": 9.887,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987081.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507987091.321, "dur": 2.912,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987091.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507987094.361, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987094.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507987095.546, "dur": 15.135,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987095.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507987110.777, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987110.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987111.993, "dur": 13.920,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987111.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507987126.009, "dur": 2.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987126.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987128.378, "dur": 3.647,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987128.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507987132.153, "dur": 0.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987132.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507987133.241, "dur": 160.353,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987133.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507987293.785, "dur": 6.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987293.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507987300.857, "dur": 4.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987300.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507987305.914, "dur": 1.535,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987305.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507987307.545, "dur": 12.064,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987307.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507987319.738, "dur": 7.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987319.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507987326.873, "dur": 1.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987326.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507987328.697, "dur": 7.584,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987328.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507987336.410, "dur": 8.959,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987336.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507987345.498, "dur": 5.823,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987345.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987351.450, "dur": 22.335,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987351.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507987373.914, "dur": 3.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987373.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987377.177, "dur": 26.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987377.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507987404.250, "dur": 1.503,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987404.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507987405.881, "dur": 28.320,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987405.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507987434.330, "dur": 6.879,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987434.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507987441.305, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987441.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987442.681, "dur": 23.201,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987442.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507987466.042, "dur": 3.007,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987466.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987469.178, "dur": 30.239,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987469.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507987499.545, "dur": 1.601,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987499.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507987501.273, "dur": 32.032,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987501.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507987535.833, "dur": 1.152,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987535.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987541.881, "dur": 51.233,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987541.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987593.241, "dur": 48.545,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987593.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507987641.913, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987641.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507987644.122, "dur": 21.119,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987644.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507987665.337, "dur": 6.625,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987665.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507987672.089, "dur": 1.249,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987672.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987673.434, "dur": 13.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987673.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507987686.841, "dur": 2.816,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987686.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987689.785, "dur": 16.576,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987689.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507987706.490, "dur": 1.247,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987706.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507987707.833, "dur": 8.961,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987707.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507987716.921, "dur": 3.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987716.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507987720.090, "dur": 1.023,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987720.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507987721.273, "dur": 15.776,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987721.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507987737.146, "dur": 1.023,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987737.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987738.298, "dur": 13.567,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987738.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507987751.993, "dur": 2.241,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987751.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987754.329, "dur": 3.616,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987754.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507987758.074, "dur": 0.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987758.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507987759.161, "dur": 153.952,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987759.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507987913.242, "dur": 7.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987913.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507987920.441, "dur": 4.896,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987920.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507987925.466, "dur": 1.503,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987925.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507987927.098, "dur": 12.095,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987927.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507987939.322, "dur": 6.976,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987939.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507987946.425, "dur": 1.729,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987946.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507987948.250, "dur": 7.583,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987948.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507987955.929, "dur": 8.801,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987955.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507987964.857, "dur": 5.792,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987964.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987970.745, "dur": 22.465,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987970.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507987993.306, "dur": 3.007,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987993.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507987996.442, "dur": 27.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507987996.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507988023.674, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988023.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507988025.305, "dur": 25.889,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988025.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507988051.322, "dur": 6.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988051.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507988058.297, "dur": 1.249,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988058.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988059.673, "dur": 23.873,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988059.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507988083.674, "dur": 3.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988083.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988086.906, "dur": 30.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988086.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507988117.273, "dur": 1.569,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988117.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507988118.969, "dur": 32.545,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988118.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507988154.457, "dur": 1.185,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.12151898734177215
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988154.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988160.569, "dur": 50.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988160.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988211.450, "dur": 48.704,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988211.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507988260.250, "dur": 2.112,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988260.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507988262.490, "dur": 21.824,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988262.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507988284.410, "dur": 6.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988284.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507988291.290, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988291.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988292.666, "dur": 13.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988292.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507988305.978, "dur": 2.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988305.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988309.049, "dur": 16.705,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988309.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507988325.850, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988325.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507988327.258, "dur": 9.376,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988327.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507988336.794, "dur": 3.551,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988336.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507988340.474, "dur": 1.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988340.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507988341.658, "dur": 14.783,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988341.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507988356.570, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988356.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988357.722, "dur": 13.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988357.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507988371.514, "dur": 2.239,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988371.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988373.882, "dur": 3.680,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988373.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507988377.689, "dur": 0.961,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988377.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507988378.778, "dur": 155.360,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988378.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507988534.330, "dur": 7.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988534.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507988541.562, "dur": 4.896,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988541.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507988546.554, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988546.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507988548.186, "dur": 11.615,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988548.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507988559.930, "dur": 7.359,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988559.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507988567.386, "dur": 1.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988567.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507988569.210, "dur": 7.616,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988569.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507988576.922, "dur": 9.952,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988576.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507988586.970, "dur": 6.336,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988586.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988593.402, "dur": 22.688,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988593.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507988616.218, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988616.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988619.450, "dur": 26.976,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988619.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507988646.553, "dur": 1.505,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988646.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507988648.186, "dur": 24.864,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988648.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507988673.210, "dur": 6.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988673.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507988680.186, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988680.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988681.562, "dur": 22.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988681.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507988704.634, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988704.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988707.866, "dur": 30.048,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988707.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507988738.042, "dur": 1.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988738.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507988739.770, "dur": 31.488,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988739.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507988773.850, "dur": 1.024,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988773.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988779.738, "dur": 51.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988779.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988830.970, "dur": 48.160,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988830.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507988879.258, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988879.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507988881.466, "dur": 21.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988881.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507988902.778, "dur": 6.560,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988902.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507988909.466, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988909.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988910.810, "dur": 13.408,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988910.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507988924.346, "dur": 2.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988924.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988927.322, "dur": 16.640,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988927.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507988944.122, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988944.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507988945.466, "dur": 9.824,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988945.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507988955.418, "dur": 2.976,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988955.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507988958.522, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988958.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507988959.706, "dur": 15.392,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988959.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507988975.226, "dur": 1.568,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988975.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988976.922, "dur": 13.568,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988976.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507988990.618, "dur": 2.208,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988990.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507988992.954, "dur": 3.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988992.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507988996.730, "dur": 0.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988996.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507988997.786, "dur": 155.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507988997.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507989153.082, "dur": 7.264,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989153.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507989160.474, "dur": 4.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989160.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507989165.530, "dur": 1.568,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989165.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507989167.194, "dur": 11.680,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989167.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507989179.002, "dur": 7.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989179.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507989186.298, "dur": 1.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989186.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507989188.090, "dur": 7.584,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989188.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507989195.802, "dur": 9.152,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989195.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507989205.082, "dur": 6.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989205.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989211.354, "dur": 23.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989211.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507989234.618, "dur": 3.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989234.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989237.914, "dur": 27.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989237.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507989265.210, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989265.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290507989266.842, "dur": 21.312,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989266.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507989288.250, "dur": 7.232,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989288.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989295.578, "dur": 1.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989295.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290507989296.954, "dur": 20.032,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989296.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290507989317.114, "dur": 13.792,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989317.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989331.034, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989331.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989332.410, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989332.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290507989333.786, "dur": 25.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989333.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989359.674, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989359.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989360.986, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989360.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290507989362.362, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989362.362,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290507989363.546, "dur": 37.856,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989363.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290507989401.530, "dur": 10.304,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989401.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989411.962, "dur": 1.920,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989411.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989414.010, "dur": 10.624,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989414.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507989424.762, "dur": 2.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989424.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989427.834, "dur": 11.872,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989427.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507989439.834, "dur": 1.152,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989439.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290507989441.114, "dur": 7.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989441.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989448.890, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989448.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290507989450.266, "dur": 4.416,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989450.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989454.842, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989454.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989456.186, "dur": 10.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989456.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507989466.458, "dur": 2.880,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989466.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989469.434, "dur": 10.624,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989469.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507989480.154, "dur": 1.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989480.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290507989481.434, "dur": 2.464,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989481.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290507989484.026, "dur": 7.424,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989484.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989491.546, "dur": 23.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989491.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507989514.682, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989514.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989517.914, "dur": 29.888,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989517.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507989547.962, "dur": 1.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989547.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507989549.690, "dur": 32.128,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989549.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507989583.962, "dur": 0.992,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989583.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989589.786, "dur": 53.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989589.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989643.546, "dur": 48.416,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989643.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507989692.090, "dur": 2.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989692.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507989694.362, "dur": 30.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989694.362,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507989724.570, "dur": 6.880,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989724.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507989731.546, "dur": 1.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989731.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989732.954, "dur": 13.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989732.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507989746.586, "dur": 2.880,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989746.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989749.594, "dur": 16.224,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989749.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507989765.946, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989765.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507989767.322, "dur": 9.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989767.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507989776.634, "dur": 2.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989776.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507989779.706, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989779.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507989780.922, "dur": 15.808,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989780.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507989796.858, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989796.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989798.010, "dur": 13.760,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989798.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507989811.898, "dur": 2.208,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989811.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507989814.234, "dur": 3.584,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989814.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507989817.946, "dur": 0.992,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989817.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507989819.066, "dur": 152.032,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989819.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507989971.226, "dur": 7.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989971.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507989978.426, "dur": 4.864,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989978.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507989983.386, "dur": 1.568,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989983.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507989985.082, "dur": 12.640,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989985.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507989997.850, "dur": 6.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507989997.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507990004.762, "dur": 1.760,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990004.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507990006.618, "dur": 7.552,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990006.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507990014.298, "dur": 8.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990014.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507990023.418, "dur": 5.984,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990023.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990029.530, "dur": 22.592,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990029.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507990052.250, "dur": 3.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990052.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990055.482, "dur": 27.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990055.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507990082.618, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990082.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507990084.282, "dur": 23.968,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990084.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507990108.378, "dur": 6.912,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990108.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507990115.418, "dur": 1.376,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990115.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990116.890, "dur": 23.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990116.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507990140.090, "dur": 3.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990140.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990143.354, "dur": 29.760,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990143.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507990173.274, "dur": 1.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990173.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507990175.066, "dur": 31.904,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990175.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507990208.602, "dur": 1.088,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990208.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990214.490, "dur": 50.656,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990214.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990265.274, "dur": 48.736,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990265.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507990314.138, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990314.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507990316.346, "dur": 20.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990316.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507990337.434, "dur": 6.720,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990337.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507990344.282, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990344.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990345.658, "dur": 13.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990345.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507990359.002, "dur": 2.912,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990359.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990362.042, "dur": 16.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990362.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507990379.098, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990379.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507990380.506, "dur": 9.312,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990380.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507990389.946, "dur": 3.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990389.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507990393.114, "dur": 1.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990393.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507990394.298, "dur": 16.000,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990394.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507990410.426, "dur": 1.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990410.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990411.738, "dur": 13.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990411.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507990425.498, "dur": 2.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990425.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990427.866, "dur": 3.776,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990427.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507990431.770, "dur": 0.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990431.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507990432.858, "dur": 153.568,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990432.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507990586.554, "dur": 6.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990586.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507990593.626, "dur": 4.832,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990593.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507990598.554, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990598.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507990600.218, "dur": 11.584,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990600.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507990611.930, "dur": 7.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990611.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507990619.162, "dur": 1.696,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990619.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507990620.954, "dur": 7.776,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990620.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507990628.858, "dur": 9.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990628.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507990638.298, "dur": 6.432,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990638.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990644.890, "dur": 23.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990644.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507990668.154, "dur": 3.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990668.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990671.386, "dur": 27.264,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990671.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507990698.778, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990698.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507990700.442, "dur": 25.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990700.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507990726.042, "dur": 7.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990726.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507990733.210, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990733.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990734.554, "dur": 23.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990734.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507990757.786, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990757.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990760.986, "dur": 29.888,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990760.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507990790.970, "dur": 1.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990790.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507990792.730, "dur": 31.584,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990792.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507990827.002, "dur": 0.992,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990827.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990832.794, "dur": 49.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990832.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990882.522, "dur": 48.448,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990882.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507990931.130, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990931.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507990933.338, "dur": 22.208,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990933.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507990955.642, "dur": 6.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990955.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507990962.522, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990962.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990963.898, "dur": 13.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990963.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507990977.242, "dur": 2.880,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990977.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507990980.218, "dur": 16.000,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990980.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507990996.346, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990996.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507990997.722, "dur": 9.408,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507990997.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507991007.258, "dur": 3.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991007.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507991010.458, "dur": 1.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991010.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507991011.642, "dur": 14.912,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991011.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507991026.682, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991026.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991027.834, "dur": 13.856,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991027.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507991041.786, "dur": 2.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991041.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991044.154, "dur": 3.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991044.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507991047.930, "dur": 0.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991047.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507991049.018, "dur": 156.096,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991049.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507991205.274, "dur": 7.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991205.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507991212.410, "dur": 4.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991212.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507991217.434, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991217.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507991219.098, "dur": 11.712,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991219.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507991230.938, "dur": 7.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991230.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507991238.106, "dur": 1.696,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991238.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507991239.930, "dur": 8.384,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991239.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507991248.442, "dur": 10.720,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991248.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507991259.290, "dur": 5.984,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991259.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991265.434, "dur": 23.232,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991265.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507991288.794, "dur": 3.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991288.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991292.090, "dur": 27.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991292.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507991319.322, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991319.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507991320.986, "dur": 23.296,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991320.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507991344.378, "dur": 6.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991344.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507991351.322, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991351.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991352.666, "dur": 22.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991352.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507991375.642, "dur": 3.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991375.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991378.810, "dur": 29.984,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991378.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507991408.922, "dur": 1.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991408.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507991410.682, "dur": 31.744,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991410.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507991445.466, "dur": 1.056,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991445.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991451.386, "dur": 52.832,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991451.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991504.378, "dur": 48.864,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991504.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507991553.370, "dur": 2.048,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991553.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507991555.546, "dur": 21.152,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991555.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507991576.826, "dur": 6.688,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991576.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507991583.642, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991583.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991585.018, "dur": 13.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991585.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507991598.330, "dur": 2.880,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991598.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991601.306, "dur": 16.864,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991601.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507991618.298, "dur": 1.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991618.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507991619.706, "dur": 9.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991619.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507991629.018, "dur": 3.520,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991629.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507991632.666, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991632.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507991633.850, "dur": 15.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991633.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507991648.986, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991648.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991650.138, "dur": 13.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991650.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507991663.898, "dur": 2.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991663.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991666.266, "dur": 3.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991666.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507991670.042, "dur": 0.960,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991670.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507991671.130, "dur": 159.616,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991671.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507991830.906, "dur": 6.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991830.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507991837.914, "dur": 4.832,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991837.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507991842.906, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991842.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507991844.538, "dur": 12.672,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991844.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507991857.338, "dur": 6.720,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991857.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507991864.186, "dur": 1.696,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991864.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507991866.010, "dur": 7.360,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991866.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507991873.498, "dur": 8.512,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991873.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507991882.138, "dur": 6.016,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991882.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991888.314, "dur": 22.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991888.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507991910.586, "dur": 3.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991910.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991913.818, "dur": 27.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991913.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507991941.018, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991941.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507991942.682, "dur": 23.616,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991942.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507991966.426, "dur": 6.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991966.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507991973.403, "dur": 1.215,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991973.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507991974.778, "dur": 23.009,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991974.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507991997.914, "dur": 3.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507991997.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992001.082, "dur": 30.112,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992001.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507992031.290, "dur": 1.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992031.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507992033.050, "dur": 31.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992033.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507992066.042, "dur": 1.152,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992066.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992072.026, "dur": 50.752,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992072.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992122.906, "dur": 48.608,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992122.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507992171.642, "dur": 2.208,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992171.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507992173.978, "dur": 21.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992173.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507992195.738, "dur": 6.656,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992195.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507992202.490, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992202.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992203.866, "dur": 13.377,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992203.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507992217.370, "dur": 2.849,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992217.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992220.346, "dur": 16.929,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992220.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507992237.370, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992237.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507992238.746, "dur": 9.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992238.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507992248.602, "dur": 3.296,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992248.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507992252.026, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992252.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507992253.211, "dur": 15.679,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992253.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507992269.018, "dur": 1.472,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992269.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992270.618, "dur": 13.792,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992270.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507992284.570, "dur": 2.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992284.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992286.938, "dur": 3.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992286.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507992290.746, "dur": 0.929,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992290.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507992291.802, "dur": 158.304,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992291.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507992450.234, "dur": 7.297,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992450.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507992457.626, "dur": 4.832,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992457.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507992462.554, "dur": 1.569,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992462.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507992464.219, "dur": 11.904,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992464.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507992476.218, "dur": 7.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992476.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507992483.386, "dur": 1.696,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992483.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507992485.211, "dur": 8.544,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992485.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507992493.882, "dur": 10.752,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992493.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507992504.763, "dur": 6.175,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992504.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992511.098, "dur": 23.137,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992511.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507992534.395, "dur": 3.103,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992534.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992537.658, "dur": 26.913,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992537.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507992564.698, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992564.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507992566.331, "dur": 22.368,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992566.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507992588.795, "dur": 6.879,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992588.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507992595.802, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992595.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992597.147, "dur": 23.039,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992597.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507992620.315, "dur": 3.039,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992620.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992623.450, "dur": 30.049,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992623.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507992653.626, "dur": 1.633,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992653.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507992655.354, "dur": 31.264,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992655.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507992689.722, "dur": 1.025,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.1404878048780488
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992689.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992695.546, "dur": 50.337,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992695.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992746.011, "dur": 48.448,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992746.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507992794.586, "dur": 2.049,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992794.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507992796.795, "dur": 21.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992796.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507992818.203, "dur": 6.655,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992818.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507992824.987, "dur": 1.247,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992824.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992826.331, "dur": 13.343,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992826.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507992839.771, "dur": 2.912,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992839.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992842.810, "dur": 16.416,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992842.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507992859.355, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992859.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507992860.730, "dur": 9.440,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992860.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507992870.299, "dur": 3.391,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992870.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507992873.819, "dur": 1.087,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992873.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507992875.035, "dur": 15.104,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992875.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507992890.266, "dur": 1.089,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992890.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992891.482, "dur": 13.729,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992891.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507992905.338, "dur": 2.240,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992905.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507992907.707, "dur": 3.647,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992907.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507992911.483, "dur": 0.927,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992911.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507992912.570, "dur": 153.121,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507992912.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507993065.882, "dur": 7.041,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993065.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507993073.019, "dur": 4.800,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993073.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507993077.946, "dur": 1.505,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993077.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507993079.579, "dur": 12.416,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993079.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507993092.091, "dur": 6.720,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993092.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507993098.938, "dur": 1.696,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993098.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507993100.794, "dur": 7.553,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993100.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507993108.442, "dur": 9.633,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993108.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507993118.203, "dur": 6.208,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993118.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993124.571, "dur": 22.528,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993124.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507993147.195, "dur": 3.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993147.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993150.491, "dur": 26.624,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993150.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507993177.242, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993177.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290507993178.907, "dur": 24.832,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993178.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507993203.835, "dur": 6.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993203.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993210.714, "dur": 1.249,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993210.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993212.059, "dur": 23.135,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993212.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507993235.323, "dur": 3.167,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993235.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993238.587, "dur": 30.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993238.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507993268.859, "dur": 1.600,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993268.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507993270.587, "dur": 32.000,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993270.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507993304.091, "dur": 1.120,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993304.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993309.979, "dur": 53.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993309.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993363.163, "dur": 48.480,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993363.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507993411.771, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993411.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290507993414.011, "dur": 21.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993414.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290507993435.195, "dur": 6.752,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993435.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993442.075, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993442.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993443.450, "dur": 13.377,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993443.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507993456.955, "dur": 2.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993456.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993459.963, "dur": 16.864,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993459.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507993476.923, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993476.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507993478.298, "dur": 9.217,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993478.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507993487.611, "dur": 2.912,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993487.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507993490.619, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993490.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507993491.803, "dur": 15.424,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993491.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507993507.355, "dur": 1.632,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993507.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993509.147, "dur": 13.568,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993509.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507993522.843, "dur": 2.207,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993522.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993525.179, "dur": 3.584,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993525.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507993528.923, "dur": 0.896,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993528.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507993529.979, "dur": 159.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993529.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507993689.787, "dur": 7.008,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993689.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507993696.955, "dur": 5.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993696.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507993702.139, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993702.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507993703.771, "dur": 11.744,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993703.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507993715.643, "dur": 7.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993715.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507993722.779, "dur": 1.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993722.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507993724.635, "dur": 7.616,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993724.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507993732.379, "dur": 9.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993732.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507993741.723, "dur": 6.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993741.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993747.995, "dur": 23.296,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993747.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507993771.419, "dur": 3.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993771.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993774.683, "dur": 27.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993774.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507993801.851, "dur": 1.504,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993801.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290507993803.483, "dur": 23.168,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993803.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290507993826.779, "dur": 7.040,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993826.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993833.979, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993833.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290507993835.323, "dur": 19.232,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993835.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290507993854.651, "dur": 13.568,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993854.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993868.347, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993868.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993869.723, "dur": 1.248,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993869.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290507993871.099, "dur": 25.792,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993871.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993896.987, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993896.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993898.331, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993898.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290507993899.675, "dur": 1.024,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993899.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290507993900.859, "dur": 37.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993900.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290507993938.267, "dur": 12.608,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993938.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993950.843, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993950.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993952.187, "dur": 10.784,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993952.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507993963.067, "dur": 2.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993963.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993966.107, "dur": 12.544,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993966.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507993978.779, "dur": 1.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993978.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290507993980.059, "dur": 7.648,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993980.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993987.803, "dur": 1.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993987.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290507993989.179, "dur": 4.512,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993989.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507993993.787, "dur": 1.216,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993993.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507993995.131, "dur": 10.304,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507993995.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507994005.563, "dur": 2.848,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994005.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994008.539, "dur": 10.560,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994008.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290507994019.227, "dur": 1.184,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994019.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290507994020.507, "dur": 2.560,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994020.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290507994023.195, "dur": 8.000,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994023.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994031.323, "dur": 23.072,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994031.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507994054.523, "dur": 3.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994054.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994057.787, "dur": 29.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994057.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507994087.643, "dur": 1.664,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994087.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290507994089.435, "dur": 31.904,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994089.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290507994124.251, "dur": 1.056,
    "args": {
      "External id": 1290,      "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994124.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994130.171, "dur": 54.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994130.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994185.243, "dur": 48.768,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994185.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290507994234.139, "dur": 2.080,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994234.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290507994236.379, "dur": 32.320,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994236.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994268.827, "dur": 17.792,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994268.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290507994286.747, "dur": 10.176,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994286.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290507994297.051, "dur": 17.056,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994297.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994314.235, "dur": 4.384,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994314.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290507994318.747, "dur": 155.456,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994318.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290507994474.331, "dur": 13.312,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994474.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290507994487.771, "dur": 7.392,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994487.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290507994495.291, "dur": 6.144,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994495.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290507994501.563, "dur": 4.896,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994501.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290507994506.587, "dur": 5.952,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994506.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994512.667, "dur": 28.032,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994512.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290507994540.795, "dur": 20.608,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994540.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507994561.531, "dur": 1.280,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994561.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994562.907, "dur": 13.952,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994562.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507994576.955, "dur": 2.944,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994576.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290507994580.027, "dur": 2.176,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994580.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290507994582.331, "dur": 3.456,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994582.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290507994585.915, "dur": 1.088,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994585.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290507994587.131, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994587.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994588.795, "dur": 13.792,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994588.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507994602.715, "dur": 2.112,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994602.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290507994604.987, "dur": 0.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994604.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290507994606.075, "dur": 4.928,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994606.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290507994611.131, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994611.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290507994612.795, "dur": 7.360,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994612.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290507994620.283, "dur": 1.728,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994620.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290507994622.107, "dur": 22.464,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994622.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290507994644.667, "dur": 3.136,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994644.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290507994647.931, "dur": 1.536,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994647.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290507994649.595, "dur": 24.768,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994649.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290507994674.491, "dur": 1.440,
    "args": {
      "External id": 1290,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 503491,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 503491, "pid": 0, "tid": 7, "ts": 1290507994674.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507965911.822, "dur": 827.793,
    "args": {
      "External id": 1290,      "cbid": 311, "correlation": 503491
    }
  },
  {
    "ph": "s", "id": 503491, "pid": 14565, "tid": 14640, "ts": 1290507965911.822,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507970393.439, "dur": 1.471,
    "args": {
      "External id": 2058,      "cbid": 317, "correlation": 504266
    }
  },
  {
    "ph": "f", "id": 504266, "pid": 14565, "tid": 14640, "ts": 1290507970393.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290507994677.691, "dur": 8.480,
    "args": {
      "External id": 2058,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504268,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504268, "pid": 0, "tid": 7, "ts": 1290507994677.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290507994686.523, "dur": 16.160,
    "args": {
      "External id": 2058,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504268,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504268, "pid": 0, "tid": 7, "ts": 1290507994686.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507970395.965, "dur": 22.343,
    "args": {
      "External id": 2058,      "cbid": 311, "correlation": 504268
    }
  },
  {
    "ph": "s", "id": 504268, "pid": 14565, "tid": 14640, "ts": 1290507970395.965,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290507994705.723, "dur": 22.336,
    "args": {
      "External id": 2060,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504279,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504279, "pid": 0, "tid": 7, "ts": 1290507994705.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507970481.382, "dur": 14.516,
    "args": {
      "External id": 2060,      "cbid": 211, "correlation": 504279
    }
  },
  {
    "ph": "s", "id": 504279, "pid": 14565, "tid": 14640, "ts": 1290507970481.382,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290508001171.196, "dur": 2.016,
    "args": {
      "External id": 2068,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504364,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504364, "pid": 0, "tid": 7, "ts": 1290508001171.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507975870.103, "dur": 12.252,
    "args": {
      "External id": 2068,      "cbid": 211, "correlation": 504364
    }
  },
  {
    "ph": "s", "id": 504364, "pid": 14565, "tid": 14640, "ts": 1290507975870.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507975890.741, "dur": 0.550,
    "args": {
      "External id": 2066,      "cbid": 317, "correlation": 504372
    }
  },
  {
    "ph": "f", "id": 504372, "pid": 14565, "tid": 14640, "ts": 1290507975890.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290508001177.436, "dur": 12.320,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001177.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290508001189.884, "dur": 18.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001189.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290508001208.028, "dur": 6.656,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001208.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508001214.780, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001214.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001216.124, "dur": 23.296,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001216.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508001239.548, "dur": 3.232,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001239.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001242.876, "dur": 30.176,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001242.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508001273.148, "dur": 1.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001273.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508001274.908, "dur": 33.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001274.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508001310.684, "dur": 1.120,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001310.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001316.700, "dur": 55.968,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001316.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001372.764, "dur": 48.704,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001372.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508001421.596, "dur": 2.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001421.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290508001423.772, "dur": 29.984,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001423.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508001453.884, "dur": 6.528,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001453.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508001460.508, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001460.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001461.884, "dur": 13.632,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001461.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508001475.644, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001475.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001478.652, "dur": 16.160,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001478.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508001494.972, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001494.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508001496.316, "dur": 9.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001496.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508001505.916, "dur": 3.104,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001505.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508001509.148, "dur": 1.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001509.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508001510.332, "dur": 15.648,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001510.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508001526.108, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001526.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001527.260, "dur": 13.888,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001527.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508001541.276, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001541.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001543.644, "dur": 3.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001543.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508001547.452, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001547.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508001548.572, "dur": 156.896,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001548.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508001705.660, "dur": 6.944,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001705.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508001712.732, "dur": 4.832,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001712.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508001717.692, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001717.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508001719.356, "dur": 12.096,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001719.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508001731.580, "dur": 7.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001731.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508001738.748, "dur": 1.696,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001738.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508001740.572, "dur": 7.520,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001740.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508001748.220, "dur": 8.800,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001748.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508001757.148, "dur": 5.824,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001757.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001763.100, "dur": 22.176,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001763.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508001785.372, "dur": 3.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001785.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001788.668, "dur": 27.264,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001788.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508001816.092, "dur": 1.472,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001816.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508001817.724, "dur": 20.736,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001817.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508001838.588, "dur": 6.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001838.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508001845.564, "dur": 1.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001845.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001846.940, "dur": 23.520,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001846.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508001870.556, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001870.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001873.820, "dur": 29.920,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001873.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508001903.868, "dur": 1.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001903.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508001905.596, "dur": 31.616,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001905.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508001939.484, "dur": 1.152,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001939.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001945.500, "dur": 52.160,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001945.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508001997.788, "dur": 47.456,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508001997.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508002045.404, "dur": 2.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002045.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508002047.612, "dur": 21.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002047.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508002069.148, "dur": 6.560,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002069.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508002075.836, "dur": 1.249,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002075.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002077.180, "dur": 13.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002077.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508002090.492, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002090.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002093.501, "dur": 16.416,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002093.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508002110.077, "dur": 1.215,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002110.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508002111.420, "dur": 9.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002111.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508002120.604, "dur": 2.816,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002120.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508002123.548, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002123.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508002124.732, "dur": 15.361,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002124.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508002140.220, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002140.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002141.340, "dur": 13.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002141.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508002154.876, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002154.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002157.212, "dur": 3.616,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002157.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508002160.988, "dur": 0.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002160.988,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508002162.076, "dur": 154.368,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002162.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508002316.604, "dur": 7.392,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002316.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508002324.124, "dur": 4.833,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002324.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508002329.084, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002329.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508002330.716, "dur": 11.744,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002330.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508002342.588, "dur": 7.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002342.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508002349.788, "dur": 1.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002349.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508002351.581, "dur": 7.359,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002351.581,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508002359.069, "dur": 9.663,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002359.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508002368.860, "dur": 6.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002368.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002375.036, "dur": 21.952,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002375.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508002397.084, "dur": 3.105,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002397.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002400.316, "dur": 26.720,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002400.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508002427.164, "dur": 1.472,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002427.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508002428.765, "dur": 25.184,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002428.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508002454.076, "dur": 6.688,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002454.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508002460.892, "dur": 1.249,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002460.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002462.237, "dur": 22.911,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002462.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508002485.276, "dur": 3.041,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002485.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002488.444, "dur": 30.049,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002488.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508002518.620, "dur": 1.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002518.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508002520.413, "dur": 30.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002520.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508002553.916, "dur": 0.992,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002553.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002559.773, "dur": 52.575,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002559.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002612.477, "dur": 48.319,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002612.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508002660.925, "dur": 2.047,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002660.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508002663.132, "dur": 21.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002663.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508002684.796, "dur": 6.625,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002684.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508002691.548, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002691.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002692.892, "dur": 13.345,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002692.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508002706.397, "dur": 2.847,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002706.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002709.373, "dur": 16.576,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002709.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508002726.044, "dur": 1.249,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002726.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508002727.389, "dur": 9.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002727.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508002737.084, "dur": 2.720,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002737.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508002739.964, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002739.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508002741.149, "dur": 15.007,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002741.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508002756.252, "dur": 1.729,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002756.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002758.108, "dur": 13.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002758.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508002771.805, "dur": 2.207,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002771.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002774.141, "dur": 3.583,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002774.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508002777.853, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002777.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508002778.909, "dur": 156.479,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002778.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508002935.517, "dur": 7.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002935.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508002942.684, "dur": 4.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002942.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508002947.741, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002947.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508002949.372, "dur": 11.809,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002949.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508002961.309, "dur": 7.008,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002961.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508002968.413, "dur": 1.696,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002968.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508002970.204, "dur": 7.617,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002970.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508002977.948, "dur": 8.417,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002977.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508002986.493, "dur": 6.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002986.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508002992.701, "dur": 23.104,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508002992.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508003015.901, "dur": 3.167,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003015.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003019.165, "dur": 27.007,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003019.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508003046.333, "dur": 1.471,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003046.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508003047.964, "dur": 25.473,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003047.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508003073.565, "dur": 6.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003073.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508003080.605, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003080.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003081.949, "dur": 23.071,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003081.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508003105.149, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003105.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003108.413, "dur": 30.176,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003108.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508003138.717, "dur": 1.631,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003138.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508003140.477, "dur": 32.447,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003140.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508003174.493, "dur": 1.023,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.14076246334310852
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003174.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003180.381, "dur": 50.112,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003180.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003230.589, "dur": 48.384,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003230.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508003279.133, "dur": 2.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003279.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508003281.341, "dur": 21.951,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003281.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508003303.421, "dur": 6.624,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003303.421,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508003310.140, "dur": 1.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003310.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003311.517, "dur": 13.376,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003311.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508003325.020, "dur": 2.945,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003325.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003328.061, "dur": 16.160,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003328.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508003344.349, "dur": 1.279,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003344.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508003345.757, "dur": 9.408,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003345.757,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508003355.325, "dur": 2.943,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003355.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508003358.428, "dur": 1.025,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003358.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508003359.613, "dur": 14.656,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003359.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508003374.397, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003374.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003375.549, "dur": 13.408,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003375.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508003389.084, "dur": 2.209,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003389.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003391.420, "dur": 3.745,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003391.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508003395.293, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003395.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508003396.381, "dur": 160.032,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003396.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508003556.605, "dur": 6.943,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003556.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508003563.709, "dur": 4.864,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003563.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508003568.669, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003568.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508003570.301, "dur": 11.776,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003570.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508003582.205, "dur": 7.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003582.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508003589.373, "dur": 1.696,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003589.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508003591.197, "dur": 7.776,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003591.197,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508003599.101, "dur": 10.112,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003599.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508003609.309, "dur": 6.144,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003609.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003615.549, "dur": 22.688,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003615.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508003638.397, "dur": 3.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003638.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003641.597, "dur": 27.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003641.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508003668.861, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003668.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508003670.525, "dur": 22.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003670.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508003693.629, "dur": 6.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003693.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508003700.605, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003700.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003701.981, "dur": 23.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003701.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508003725.277, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003725.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003728.509, "dur": 30.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003728.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508003758.909, "dur": 1.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003758.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508003760.637, "dur": 32.224,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003760.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508003794.941, "dur": 1.024,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003794.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003800.829, "dur": 52.224,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003800.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003853.181, "dur": 48.480,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003853.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508003901.789, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003901.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508003903.997, "dur": 20.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003903.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508003925.053, "dur": 6.688,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003925.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508003931.837, "dur": 1.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003931.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003933.213, "dur": 13.120,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003933.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508003946.461, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003946.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003949.469, "dur": 16.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003949.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508003966.557, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003966.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508003967.901, "dur": 9.888,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003967.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508003977.949, "dur": 2.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003977.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508003980.989, "dur": 1.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003980.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508003982.205, "dur": 15.328,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003982.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508003997.661, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003997.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508003998.813, "dur": 13.824,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508003998.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508004012.733, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004012.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004015.101, "dur": 3.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004015.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508004018.909, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004018.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508004019.997, "dur": 163.840,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004019.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508004183.965, "dur": 7.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004183.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508004191.165, "dur": 4.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004191.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508004196.221, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004196.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508004197.885, "dur": 11.584,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004197.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508004209.565, "dur": 7.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004209.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508004216.861, "dur": 1.696,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004216.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508004218.685, "dur": 7.648,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004218.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508004226.493, "dur": 9.120,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004226.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508004235.741, "dur": 6.272,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004235.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004242.109, "dur": 23.552,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004242.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508004265.789, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004265.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004269.021, "dur": 27.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004269.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508004296.221, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004296.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508004297.885, "dur": 28.288,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004297.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508004326.269, "dur": 6.784,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004326.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508004333.181, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004333.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004334.557, "dur": 23.232,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004334.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508004357.917, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004357.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004361.085, "dur": 30.016,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004361.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508004391.229, "dur": 1.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004391.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508004392.957, "dur": 32.160,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004392.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508004427.741, "dur": 1.184,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.12162162162162163
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004427.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004433.853, "dur": 50.592,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004433.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004484.541, "dur": 48.608,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004484.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508004533.277, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004533.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508004535.485, "dur": 20.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004535.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508004556.541, "dur": 6.624,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004556.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508004563.293, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004563.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004564.669, "dur": 13.312,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004564.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508004578.077, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004578.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004581.085, "dur": 16.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004581.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508004598.173, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004598.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508004599.549, "dur": 9.408,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004599.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508004609.085, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004609.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508004612.349, "dur": 1.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004612.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508004613.565, "dur": 15.424,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004613.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508004629.085, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004629.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004630.237, "dur": 13.792,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004630.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508004644.157, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004644.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004646.525, "dur": 3.648,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004646.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508004650.333, "dur": 0.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004650.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508004651.389, "dur": 169.856,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004651.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508004821.437, "dur": 7.104,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004821.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508004828.669, "dur": 4.864,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004828.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508004833.693, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004833.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508004835.325, "dur": 12.256,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004835.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508004847.709, "dur": 7.008,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004847.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508004854.845, "dur": 1.728,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004854.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508004856.669, "dur": 7.552,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004856.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508004864.317, "dur": 9.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004864.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508004873.437, "dur": 5.920,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004873.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004879.453, "dur": 22.368,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004879.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508004901.949, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004901.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004905.149, "dur": 27.360,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004905.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508004932.605, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004932.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508004934.269, "dur": 25.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004934.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508004959.485, "dur": 6.784,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004959.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508004966.397, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004966.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004967.773, "dur": 23.264,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004967.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508004991.165, "dur": 3.008,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004991.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508004994.301, "dur": 29.856,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508004994.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508005024.285, "dur": 1.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005024.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508005025.981, "dur": 32.000,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005025.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508005059.581, "dur": 0.992,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005059.581,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005065.469, "dur": 51.968,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005065.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005117.565, "dur": 48.384,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005117.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508005166.077, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005166.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508005168.285, "dur": 21.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005168.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508005189.949, "dur": 6.720,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005189.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508005196.765, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005196.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005198.141, "dur": 13.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005198.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508005211.741, "dur": 2.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005211.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005214.717, "dur": 16.864,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005214.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508005231.709, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005231.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508005233.085, "dur": 9.408,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005233.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508005242.621, "dur": 3.616,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005242.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508005246.365, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005246.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508005247.549, "dur": 14.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005247.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508005262.589, "dur": 0.992,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005262.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005263.709, "dur": 13.472,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005263.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508005277.309, "dur": 2.208,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005277.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005279.645, "dur": 3.616,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005279.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508005283.389, "dur": 0.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005283.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508005284.445, "dur": 151.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005284.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508005435.741, "dur": 7.104,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005435.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508005442.973, "dur": 4.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005442.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508005448.029, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005448.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508005449.661, "dur": 11.648,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005449.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508005461.405, "dur": 7.296,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005461.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508005468.797, "dur": 1.728,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005468.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508005470.621, "dur": 7.808,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005470.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508005478.525, "dur": 9.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005478.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508005488.285, "dur": 6.016,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005488.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005494.429, "dur": 22.208,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005494.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508005516.765, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005516.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005519.997, "dur": 26.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005519.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508005547.101, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005547.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508005548.765, "dur": 24.512,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005548.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508005573.405, "dur": 6.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005573.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508005580.445, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005580.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005581.821, "dur": 23.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005581.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508005604.989, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005604.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005608.189, "dur": 30.112,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005608.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508005638.429, "dur": 1.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005638.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508005640.157, "dur": 32.032,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005640.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508005674.109, "dur": 0.992,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005674.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005679.965, "dur": 51.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005679.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005731.229, "dur": 48.544,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005731.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508005779.901, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005779.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508005782.109, "dur": 22.016,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005782.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508005804.221, "dur": 6.784,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005804.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508005811.133, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005811.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005812.477, "dur": 13.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005812.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508005826.045, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005826.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005829.053, "dur": 16.480,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005829.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508005845.661, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005845.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508005847.037, "dur": 9.728,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005847.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508005856.893, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005856.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508005860.093, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005860.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508005861.277, "dur": 15.584,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005861.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508005876.989, "dur": 1.120,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005876.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005878.237, "dur": 13.824,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005878.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508005892.189, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005892.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508005894.557, "dur": 3.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005894.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508005898.365, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005898.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508005899.453, "dur": 162.720,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508005899.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508006062.301, "dur": 7.392,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006062.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508006069.821, "dur": 4.832,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006069.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508006074.781, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006074.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508006076.413, "dur": 11.648,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006076.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508006088.189, "dur": 7.008,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006088.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508006095.325, "dur": 1.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006095.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508006097.149, "dur": 7.840,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006097.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508006105.117, "dur": 9.472,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006105.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508006114.749, "dur": 6.112,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006114.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006120.989, "dur": 23.488,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006120.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508006144.605, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006144.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006147.869, "dur": 27.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006147.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508006175.069, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006175.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508006176.701, "dur": 20.768,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006176.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508006197.565, "dur": 7.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006197.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006204.797, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006204.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508006206.173, "dur": 19.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006206.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508006225.309, "dur": 13.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006225.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006238.845, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006238.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006240.221, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006240.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508006241.597, "dur": 25.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006241.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006267.229, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006267.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006268.573, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006268.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508006269.917, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006269.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508006271.101, "dur": 37.472,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006271.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508006308.701, "dur": 10.656,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006308.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006319.485, "dur": 1.920,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006319.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006321.533, "dur": 10.560,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006321.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508006332.221, "dur": 2.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006332.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006335.325, "dur": 12.256,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006335.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508006347.741, "dur": 1.120,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006347.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508006349.021, "dur": 7.552,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006349.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006356.669, "dur": 1.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006356.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508006358.045, "dur": 4.448,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006358.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006362.621, "dur": 1.184,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006362.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006363.965, "dur": 10.144,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006363.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508006374.237, "dur": 2.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006374.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006377.213, "dur": 10.592,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006377.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508006387.901, "dur": 1.184,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006387.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290508006389.213, "dur": 2.432,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006389.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290508006391.773, "dur": 7.392,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006391.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006399.293, "dur": 23.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006399.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508006423.101, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006423.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006426.301, "dur": 29.952,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006426.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508006456.381, "dur": 1.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006456.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508006458.077, "dur": 32.480,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006458.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508006492.253, "dur": 1.024,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006492.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006498.173, "dur": 52.864,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006498.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006551.165, "dur": 48.512,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006551.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508006599.805, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006599.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508006602.013, "dur": 29.952,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006602.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508006632.061, "dur": 6.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006632.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508006639.005, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006639.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006640.381, "dur": 13.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006640.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508006653.917, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006653.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006656.893, "dur": 16.160,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006656.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508006673.181, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006673.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508006674.557, "dur": 8.896,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006674.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508006683.581, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006683.581,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508006686.621, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006686.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508006687.805, "dur": 15.008,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006687.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508006702.941, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006702.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006704.285, "dur": 13.792,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006704.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508006718.205, "dur": 2.208,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006718.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006720.541, "dur": 3.616,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006720.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508006724.285, "dur": 0.992,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006724.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508006725.405, "dur": 155.392,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006725.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508006880.989, "dur": 6.944,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006880.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508006888.029, "dur": 4.864,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006888.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508006893.021, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006893.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508006894.653, "dur": 12.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006894.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508006907.741, "dur": 6.816,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006907.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508006914.685, "dur": 1.792,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006914.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508006916.605, "dur": 7.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006916.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508006924.413, "dur": 8.736,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006924.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508006933.309, "dur": 6.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006933.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006939.549, "dur": 22.944,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006939.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508006962.621, "dur": 3.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006962.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508006965.917, "dur": 27.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006965.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508006993.213, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006993.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508006994.845, "dur": 23.488,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508006994.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508007018.461, "dur": 6.944,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007018.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508007025.533, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007025.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007026.909, "dur": 23.424,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007026.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508007050.429, "dur": 3.169,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007050.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007053.725, "dur": 29.760,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007053.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508007083.614, "dur": 1.631,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007083.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508007085.405, "dur": 32.064,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007085.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508007118.877, "dur": 0.992,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007118.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007124.701, "dur": 51.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007124.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007176.509, "dur": 48.480,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007176.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508007225.149, "dur": 2.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007225.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508007227.325, "dur": 21.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007227.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508007248.510, "dur": 6.783,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007248.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508007255.389, "dur": 1.249,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007255.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007256.765, "dur": 13.313,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007256.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508007270.205, "dur": 2.849,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007270.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007273.181, "dur": 16.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007273.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508007290.270, "dur": 1.247,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007290.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508007291.646, "dur": 9.855,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007291.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508007301.629, "dur": 3.520,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007301.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508007305.277, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007305.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508007306.493, "dur": 15.616,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007306.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508007322.269, "dur": 1.569,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007322.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007323.965, "dur": 13.793,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007323.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508007337.853, "dur": 2.305,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007337.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007340.254, "dur": 3.711,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007340.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508007344.093, "dur": 0.993,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007344.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508007345.181, "dur": 156.800,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007345.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508007502.141, "dur": 6.945,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007502.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508007509.213, "dur": 4.865,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007509.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508007514.173, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007514.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508007515.838, "dur": 11.551,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007515.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508007527.518, "dur": 6.879,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007527.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508007534.525, "dur": 1.696,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007534.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508007536.317, "dur": 7.777,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007536.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508007544.254, "dur": 9.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007544.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508007553.949, "dur": 6.336,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007553.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007560.414, "dur": 22.751,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007560.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508007583.293, "dur": 3.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007583.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007586.557, "dur": 27.137,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007586.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508007613.790, "dur": 1.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007613.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508007615.517, "dur": 25.697,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007615.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508007641.310, "dur": 7.039,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007641.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508007648.446, "dur": 1.247,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007648.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007649.822, "dur": 23.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007649.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508007673.597, "dur": 3.201,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007673.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007676.862, "dur": 29.759,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007676.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508007706.749, "dur": 1.632,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007706.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508007708.510, "dur": 30.783,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007708.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508007742.397, "dur": 0.992,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007742.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007748.158, "dur": 50.144,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007748.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007798.462, "dur": 47.904,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007798.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508007846.526, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007846.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508007848.733, "dur": 22.145,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007848.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508007870.974, "dur": 6.816,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007870.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508007877.886, "dur": 1.247,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007877.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007879.262, "dur": 13.247,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007879.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508007892.638, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007892.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007895.614, "dur": 15.776,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007895.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508007911.518, "dur": 1.247,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007911.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508007912.861, "dur": 9.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007912.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508007922.238, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007922.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508007925.501, "dur": 1.089,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007925.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508007926.685, "dur": 15.041,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007926.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508007941.821, "dur": 1.025,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007941.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007942.974, "dur": 13.503,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007942.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508007956.606, "dur": 2.176,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007956.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508007958.910, "dur": 3.583,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007958.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508007962.654, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007962.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508007963.709, "dur": 153.792,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508007963.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508008117.630, "dur": 7.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008117.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508008124.829, "dur": 4.865,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008124.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508008129.821, "dur": 1.505,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008129.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508008131.486, "dur": 11.679,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008131.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508008143.294, "dur": 7.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008143.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508008150.590, "dur": 1.696,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008150.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508008152.414, "dur": 8.160,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008152.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508008160.702, "dur": 10.528,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008160.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508008171.390, "dur": 6.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008171.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008177.566, "dur": 22.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008177.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508008200.574, "dur": 3.008,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008200.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008203.710, "dur": 27.168,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008203.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508008231.006, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008231.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508008232.670, "dur": 22.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008232.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508008255.646, "dur": 6.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008255.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508008262.589, "dur": 1.281,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008262.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008263.965, "dur": 21.825,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008263.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508008285.918, "dur": 3.071,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008285.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008289.118, "dur": 29.920,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008289.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508008319.165, "dur": 1.633,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008319.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508008320.926, "dur": 32.224,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008320.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508008354.813, "dur": 1.089,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.1322314049586777
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008354.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008360.734, "dur": 52.000,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008360.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008412.862, "dur": 48.128,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008412.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508008461.150, "dur": 2.016,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008461.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508008463.326, "dur": 21.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008463.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508008485.021, "dur": 6.561,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008485.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508008491.710, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008491.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008493.054, "dur": 13.120,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008493.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508008506.269, "dur": 2.881,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008506.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008509.245, "dur": 16.513,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008509.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508008525.854, "dur": 1.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008525.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508008527.230, "dur": 9.055,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008527.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508008536.414, "dur": 2.944,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008536.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508008539.486, "dur": 1.120,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008539.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508008540.734, "dur": 14.623,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008540.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508008555.454, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008555.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008556.606, "dur": 13.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008556.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508008570.302, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008570.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008572.670, "dur": 3.584,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008572.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508008576.382, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008576.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508008577.438, "dur": 156.480,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008577.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508008734.046, "dur": 6.944,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008734.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508008741.118, "dur": 4.864,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008741.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508008746.110, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008746.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508008747.742, "dur": 13.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008747.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508008760.926, "dur": 6.784,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008760.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508008767.838, "dur": 1.760,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008767.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508008769.694, "dur": 7.648,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008769.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508008777.470, "dur": 8.192,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008777.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508008785.758, "dur": 5.888,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008785.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008791.774, "dur": 22.592,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008791.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508008814.494, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008814.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008817.758, "dur": 26.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008817.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508008844.798, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008844.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508008846.430, "dur": 23.424,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008846.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508008869.982, "dur": 6.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008869.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508008876.990, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008876.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008878.334, "dur": 22.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008878.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508008901.406, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008901.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008904.606, "dur": 29.952,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008904.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508008934.686, "dur": 1.632,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008934.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508008936.446, "dur": 32.000,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008936.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508008971.390, "dur": 0.992,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008971.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508008977.278, "dur": 51.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508008977.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009028.478, "dur": 48.704,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009028.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508009077.310, "dur": 2.144,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009077.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508009079.614, "dur": 21.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009079.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508009101.406, "dur": 6.592,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009101.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508009108.126, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009108.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009109.502, "dur": 13.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009109.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508009122.878, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009122.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009125.854, "dur": 16.896,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009125.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508009142.878, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009142.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508009144.254, "dur": 9.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009144.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508009153.438, "dur": 3.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009153.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508009156.606, "dur": 1.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009156.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508009157.790, "dur": 16.064,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009157.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508009173.982, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009173.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009175.134, "dur": 13.856,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009175.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508009189.150, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009189.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009191.518, "dur": 3.680,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009191.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508009195.326, "dur": 0.960,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009195.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508009196.414, "dur": 158.304,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009196.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508009354.846, "dur": 7.296,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009354.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508009362.270, "dur": 4.800,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009362.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508009367.166, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009367.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508009368.798, "dur": 11.584,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009368.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508009380.478, "dur": 6.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009380.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508009387.454, "dur": 1.824,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009387.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508009389.406, "dur": 8.416,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009389.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508009397.950, "dur": 11.264,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009397.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508009409.342, "dur": 6.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009409.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009415.550, "dur": 23.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009415.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508009438.814, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009438.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009442.110, "dur": 26.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009442.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508009469.246, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009469.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508009470.878, "dur": 22.304,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009470.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508009493.310, "dur": 6.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009493.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508009500.414, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009500.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009501.790, "dur": 23.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009501.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508009524.958, "dur": 3.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009524.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009528.094, "dur": 30.112,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009528.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508009558.302, "dur": 1.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009558.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508009559.998, "dur": 31.552,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009559.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508009593.822, "dur": 1.088,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009593.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009599.710, "dur": 51.904,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009599.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009651.742, "dur": 48.512,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009651.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508009700.382, "dur": 2.016,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009700.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508009702.558, "dur": 21.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009702.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508009723.710, "dur": 6.656,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009723.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508009730.494, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009730.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009731.870, "dur": 13.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009731.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508009745.246, "dur": 2.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009745.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009748.286, "dur": 17.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009748.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508009765.438, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009765.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508009766.814, "dur": 9.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009766.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508009776.222, "dur": 3.200,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009776.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508009779.550, "dur": 1.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009779.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508009780.766, "dur": 15.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009780.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508009795.902, "dur": 1.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009795.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009797.118, "dur": 13.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009797.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508009810.750, "dur": 2.240,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009810.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508009813.118, "dur": 3.584,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009813.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508009816.830, "dur": 0.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009816.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508009817.886, "dur": 156.128,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009817.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508009974.206, "dur": 7.296,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009974.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508009981.598, "dur": 4.832,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009981.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508009986.526, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009986.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508009988.158, "dur": 12.160,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508009988.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508010000.446, "dur": 6.720,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010000.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508010007.326, "dur": 1.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010007.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508010009.150, "dur": 7.584,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010009.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508010016.862, "dur": 9.824,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010016.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508010026.814, "dur": 5.952,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010026.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010032.926, "dur": 22.624,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010032.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010055.678, "dur": 3.136,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010055.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010058.942, "dur": 26.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010058.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508010085.950, "dur": 1.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010085.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508010087.614, "dur": 24.352,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010087.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508010112.094, "dur": 6.784,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010112.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010119.006, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010119.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010120.350, "dur": 22.720,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010120.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010143.198, "dur": 3.040,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010143.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010146.366, "dur": 30.336,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010146.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508010176.830, "dur": 1.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010176.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508010178.558, "dur": 32.224,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010178.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508010212.350, "dur": 1.088,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010212.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010218.270, "dur": 51.008,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010218.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010269.374, "dur": 48.448,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010269.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508010317.950, "dur": 2.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010317.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508010320.126, "dur": 20.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010320.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508010341.182, "dur": 6.688,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010341.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010347.998, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010347.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010349.342, "dur": 13.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010349.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010362.686, "dur": 2.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010362.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010365.726, "dur": 16.800,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010365.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508010382.654, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010382.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508010384.030, "dur": 9.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010384.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508010393.438, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010393.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508010396.606, "dur": 1.056,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010396.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508010397.790, "dur": 15.232,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010397.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508010413.150, "dur": 1.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010413.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010414.942, "dur": 13.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010414.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010428.702, "dur": 2.176,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010428.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010431.038, "dur": 3.584,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010431.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508010434.750, "dur": 0.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010434.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508010435.838, "dur": 156.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010435.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508010593.086, "dur": 6.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010593.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508010600.062, "dur": 4.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010600.062,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508010605.118, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010605.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508010606.750, "dur": 11.968,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010606.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508010618.846, "dur": 7.232,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010618.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508010626.174, "dur": 1.728,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010626.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508010627.998, "dur": 7.456,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010627.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508010635.582, "dur": 8.992,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010635.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508010644.702, "dur": 5.984,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010644.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010650.846, "dur": 23.648,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010650.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010674.622, "dur": 3.200,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010674.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010677.950, "dur": 26.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010677.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508010705.022, "dur": 1.504,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010705.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508010706.654, "dur": 22.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010706.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508010729.630, "dur": 6.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010729.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010736.734, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010736.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508010738.110, "dur": 18.688,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010738.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508010756.894, "dur": 13.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010756.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010770.462, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010770.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010771.838, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010771.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508010773.214, "dur": 25.664,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010773.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010799.006, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010799.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010800.318, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010800.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508010801.694, "dur": 1.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010801.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508010802.846, "dur": 38.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010802.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508010841.022, "dur": 12.800,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010841.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010853.950, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010853.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010855.326, "dur": 10.816,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010855.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010866.238, "dur": 2.912,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010866.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010869.246, "dur": 11.552,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010869.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508010880.926, "dur": 1.152,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010880.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508010882.174, "dur": 7.520,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010882.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010889.790, "dur": 1.248,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010889.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508010891.134, "dur": 4.320,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010891.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508010895.550, "dur": 1.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010895.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010896.894, "dur": 10.208,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010896.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010907.230, "dur": 2.880,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010907.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010910.238, "dur": 10.560,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010910.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508010920.958, "dur": 1.120,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010920.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290508010922.206, "dur": 2.528,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010922.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290508010924.862, "dur": 7.744,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010924.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010932.702, "dur": 23.104,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010932.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508010955.934, "dur": 3.072,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010955.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508010959.134, "dur": 29.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010959.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508010988.702, "dur": 1.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010988.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508010990.398, "dur": 32.128,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508010990.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508011024.318, "dur": 1.024,
    "args": {
      "External id": 2066,      "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011024.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011030.142, "dur": 54.976,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011030.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011085.246, "dur": 48.096,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011085.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508011133.470, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011133.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290508011135.678, "dur": 33.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011135.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011169.406, "dur": 16.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011169.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508011186.462, "dur": 9.216,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011186.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508011195.806, "dur": 15.360,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011195.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011211.294, "dur": 5.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011211.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508011216.510, "dur": 154.592,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011216.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508011371.230, "dur": 13.600,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011371.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508011384.958, "dur": 7.392,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011384.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508011392.478, "dur": 6.112,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011392.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508011398.718, "dur": 4.896,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011398.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508011403.742, "dur": 5.728,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011403.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011409.598, "dur": 28.192,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011409.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290508011437.918, "dur": 20.768,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011437.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508011458.814, "dur": 1.280,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011458.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011460.222, "dur": 14.048,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011460.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508011474.398, "dur": 2.848,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011474.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508011477.374, "dur": 1.984,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011477.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508011479.486, "dur": 3.264,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011479.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508011482.878, "dur": 1.088,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011482.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508011484.094, "dur": 1.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011484.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011485.662, "dur": 13.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011485.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508011499.390, "dur": 2.080,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011499.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508011501.630, "dur": 0.928,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011501.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508011502.686, "dur": 4.896,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011502.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508011507.678, "dur": 1.568,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011507.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508011509.374, "dur": 7.424,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011509.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508011516.926, "dur": 1.728,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011516.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508011518.750, "dur": 22.400,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011518.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508011541.246, "dur": 3.104,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011541.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508011544.478, "dur": 1.536,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011544.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290508011546.142, "dur": 25.024,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011546.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508011571.294, "dur": 1.440,
    "args": {
      "External id": 2066,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 504374,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 504374, "pid": 0, "tid": 7, "ts": 1290508011571.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507975892.057, "dur": 1101.257,
    "args": {
      "External id": 2066,      "cbid": 311, "correlation": 504374
    }
  },
  {
    "ph": "s", "id": 504374, "pid": 14565, "tid": 14640, "ts": 1290507975892.057,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507980565.867, "dur": 1.320,
    "args": {
      "External id": 2834,      "cbid": 317, "correlation": 505149
    }
  },
  {
    "ph": "f", "id": 505149, "pid": 14565, "tid": 14640, "ts": 1290507980565.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290508011575.774, "dur": 8.576,
    "args": {
      "External id": 2834,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505151,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505151, "pid": 0, "tid": 7, "ts": 1290508011575.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290508011584.702, "dur": 15.936,
    "args": {
      "External id": 2834,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505151,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505151, "pid": 0, "tid": 7, "ts": 1290508011584.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507980568.149, "dur": 19.095,
    "args": {
      "External id": 2834,      "cbid": 311, "correlation": 505151
    }
  },
  {
    "ph": "s", "id": 505151, "pid": 14565, "tid": 14640, "ts": 1290507980568.149,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290508011601.726, "dur": 22.368,
    "args": {
      "External id": 2836,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505162,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505162, "pid": 0, "tid": 7, "ts": 1290508011601.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507980641.967, "dur": 12.922,
    "args": {
      "External id": 2836,      "cbid": 211, "correlation": 505162
    }
  },
  {
    "ph": "s", "id": 505162, "pid": 14565, "tid": 14640, "ts": 1290507980641.967,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290508018073.376, "dur": 1.983,
    "args": {
      "External id": 2844,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505247,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505247, "pid": 0, "tid": 7, "ts": 1290508018073.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507984818.739, "dur": 17.446,
    "args": {
      "External id": 2844,      "cbid": 211, "correlation": 505247
    }
  },
  {
    "ph": "s", "id": 505247, "pid": 14565, "tid": 14640, "ts": 1290507984818.739,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507984845.153, "dur": 0.822,
    "args": {
      "External id": 2842,      "cbid": 317, "correlation": 505255
    }
  },
  {
    "ph": "f", "id": 505255, "pid": 14565, "tid": 14640, "ts": 1290507984845.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290508018079.584, "dur": 12.256,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018079.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290508018091.968, "dur": 17.759,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018091.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290508018109.856, "dur": 6.592,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018109.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508018116.544, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018116.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018117.920, "dur": 23.423,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018117.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508018141.439, "dur": 3.233,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018141.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018144.768, "dur": 30.656,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018144.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508018175.551, "dur": 1.633,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018175.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508018177.312, "dur": 32.064,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018177.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508018210.848, "dur": 1.120,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018210.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018216.896, "dur": 56.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018216.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018273.984, "dur": 48.640,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018273.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508018322.720, "dur": 2.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018322.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290508018324.928, "dur": 29.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018324.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508018354.720, "dur": 6.559,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018354.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508018361.408, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018361.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018362.752, "dur": 13.376,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018362.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508018376.255, "dur": 2.849,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018376.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018379.231, "dur": 16.161,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018379.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508018395.520, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018395.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508018396.896, "dur": 9.408,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018396.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508018406.431, "dur": 3.521,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018406.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508018410.080, "dur": 1.088,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018410.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508018411.263, "dur": 15.073,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018411.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508018426.432, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018426.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018427.616, "dur": 13.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018427.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508018441.439, "dur": 2.273,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018441.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018443.840, "dur": 3.680,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018443.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508018447.648, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018447.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508018448.736, "dur": 171.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018448.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508018620.032, "dur": 7.008,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018620.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508018627.200, "dur": 4.864,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018627.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508018632.192, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018632.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508018633.856, "dur": 11.616,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018633.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508018645.600, "dur": 6.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018645.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508018652.672, "dur": 1.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018652.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508018654.496, "dur": 7.392,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018654.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508018661.984, "dur": 8.544,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018661.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508018670.656, "dur": 5.792,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018670.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018676.576, "dur": 22.496,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018676.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508018699.200, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018699.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018702.432, "dur": 27.296,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018702.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508018729.856, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018729.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508018731.488, "dur": 20.512,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018731.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508018752.128, "dur": 6.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018752.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508018759.040, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018759.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018760.384, "dur": 23.296,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018760.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508018783.776, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018783.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018786.976, "dur": 30.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018786.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508018817.152, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018817.152,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508018818.848, "dur": 31.488,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018818.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508018851.936, "dur": 1.120,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018851.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018857.824, "dur": 52.096,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018857.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018910.048, "dur": 48.512,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018910.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508018958.720, "dur": 2.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018958.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508018960.928, "dur": 21.152,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018960.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508018982.176, "dur": 6.592,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018982.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508018988.896, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018988.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508018990.272, "dur": 13.184,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508018990.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508019003.584, "dur": 2.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019003.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019006.656, "dur": 16.736,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019006.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508019023.552, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019023.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508019024.928, "dur": 9.376,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019024.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508019034.432, "dur": 2.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019034.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508019037.440, "dur": 1.088,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019037.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508019038.624, "dur": 14.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019038.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508019053.696, "dur": 1.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019053.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019054.848, "dur": 13.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019054.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508019068.512, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019068.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019070.880, "dur": 3.680,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019070.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508019074.688, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019074.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508019075.776, "dur": 158.208,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019075.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508019234.144, "dur": 7.328,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019234.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508019241.600, "dur": 4.864,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019241.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508019246.560, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019246.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508019248.224, "dur": 12.160,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019248.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508019260.512, "dur": 7.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019260.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508019267.680, "dur": 1.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019267.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508019269.472, "dur": 7.616,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019269.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508019277.216, "dur": 10.176,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019277.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508019287.552, "dur": 6.016,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019287.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019293.664, "dur": 22.656,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019293.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508019316.448, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019316.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019319.680, "dur": 27.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019319.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508019346.944, "dur": 1.472,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019346.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508019348.576, "dur": 25.632,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019348.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508019374.336, "dur": 6.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019374.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508019381.248, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019381.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019382.624, "dur": 22.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019382.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508019405.728, "dur": 3.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019405.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019408.992, "dur": 29.984,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019408.992,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508019439.104, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019439.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508019440.832, "dur": 30.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019440.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508019474.432, "dur": 1.184,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.12162162162162163
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019474.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019480.448, "dur": 49.792,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019480.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019530.400, "dur": 48.352,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019530.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508019578.912, "dur": 2.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019578.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508019581.120, "dur": 21.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019581.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508019602.816, "dur": 6.624,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019602.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508019609.568, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019609.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019610.944, "dur": 13.408,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019610.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508019624.512, "dur": 2.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019624.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019627.488, "dur": 16.320,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019627.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508019643.904, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019643.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508019645.280, "dur": 9.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019645.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508019654.976, "dur": 3.296,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019654.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508019658.400, "dur": 1.088,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019658.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508019659.616, "dur": 15.648,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019659.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508019675.392, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019675.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019677.088, "dur": 13.952,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019677.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508019691.168, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019691.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019693.536, "dur": 3.712,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019693.536,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508019697.344, "dur": 0.992,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019697.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508019698.432, "dur": 157.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019698.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508019856.192, "dur": 7.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019856.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508019863.360, "dur": 4.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019863.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508019868.448, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019868.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508019870.080, "dur": 11.616,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019870.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508019881.792, "dur": 7.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019881.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508019889.024, "dur": 1.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019889.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508019890.848, "dur": 7.552,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019890.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508019898.496, "dur": 8.672,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019898.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508019907.296, "dur": 5.984,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019907.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019913.376, "dur": 22.624,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019913.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508019936.096, "dur": 3.200,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019936.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508019939.392, "dur": 26.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019939.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508019966.496, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019966.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508019968.128, "dur": 25.152,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019968.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508019993.440, "dur": 6.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508019993.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508020000.384, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020000.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020001.760, "dur": 22.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020001.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508020024.800, "dur": 3.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020024.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020028.064, "dur": 29.632,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020028.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508020057.824, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020057.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508020059.520, "dur": 32.096,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020059.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508020092.864, "dur": 1.152,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020092.864,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020098.848, "dur": 50.656,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020098.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020149.600, "dur": 48.256,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020149.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508020198.016, "dur": 2.112,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020198.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508020200.256, "dur": 21.280,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020200.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508020221.632, "dur": 6.624,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020221.632,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508020228.352, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020228.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020229.696, "dur": 13.408,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020229.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508020243.232, "dur": 2.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020243.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020246.208, "dur": 15.968,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020246.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508020262.304, "dur": 1.280,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020262.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508020263.712, "dur": 8.832,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020263.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508020272.672, "dur": 2.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020272.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508020275.744, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020275.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508020276.928, "dur": 15.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020276.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508020292.096, "dur": 1.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020292.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020293.216, "dur": 13.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020293.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508020307.040, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020307.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020309.408, "dur": 3.648,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020309.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508020313.184, "dur": 0.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020313.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508020314.240, "dur": 164.224,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020314.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508020478.656, "dur": 6.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020478.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508020485.792, "dur": 4.736,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020485.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508020490.624, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020490.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508020492.256, "dur": 11.552,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020492.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508020503.936, "dur": 7.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020503.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508020511.136, "dur": 1.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020511.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508020512.928, "dur": 7.744,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020512.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508020520.800, "dur": 10.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020520.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508020530.976, "dur": 6.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020530.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020537.120, "dur": 22.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020537.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508020560.192, "dur": 2.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020560.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020563.328, "dur": 27.552,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020563.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508020590.976, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020590.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508020592.672, "dur": 23.008,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020592.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508020615.808, "dur": 6.688,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020615.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508020622.624, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020622.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020623.968, "dur": 22.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020623.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508020647.072, "dur": 3.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020647.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020650.240, "dur": 29.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020650.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508020680.064, "dur": 1.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020680.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508020681.760, "dur": 31.360,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020681.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508020715.488, "dur": 1.120,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020715.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020721.472, "dur": 52.288,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020721.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020773.856, "dur": 47.424,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020773.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508020821.408, "dur": 2.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020821.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508020823.584, "dur": 20.576,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020823.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508020844.256, "dur": 6.560,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020844.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508020850.944, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020850.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020852.288, "dur": 13.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020852.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508020865.472, "dur": 2.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020865.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020868.416, "dur": 16.512,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020868.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508020885.024, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020885.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508020886.400, "dur": 9.632,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020886.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508020896.192, "dur": 2.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020896.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508020899.200, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020899.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508020900.384, "dur": 14.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020900.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508020915.488, "dur": 0.992,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020915.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020916.608, "dur": 13.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020916.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508020930.336, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020930.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508020932.704, "dur": 3.616,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020932.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508020936.448, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020936.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508020937.504, "dur": 153.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508020937.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508021090.688, "dur": 7.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021090.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508021097.984, "dur": 4.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021097.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508021103.040, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021103.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508021104.640, "dur": 11.680,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021104.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508021116.448, "dur": 6.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021116.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508021123.488, "dur": 1.664,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021123.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508021125.312, "dur": 7.392,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021125.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508021132.832, "dur": 9.184,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021132.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508021142.112, "dur": 5.952,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021142.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021148.160, "dur": 22.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021148.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508021171.072, "dur": 3.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021171.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021174.368, "dur": 26.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021174.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508021201.312, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021201.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508021202.976, "dur": 27.872,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021202.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508021230.944, "dur": 6.720,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021230.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508021237.760, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021237.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021239.136, "dur": 22.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021239.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508021262.272, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021262.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021265.504, "dur": 29.344,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021265.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508021294.976, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021294.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508021296.704, "dur": 31.808,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021296.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508021329.888, "dur": 1.088,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021329.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021335.872, "dur": 50.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021335.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021386.816, "dur": 48.704,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021386.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508021435.648, "dur": 2.144,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021435.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508021437.920, "dur": 20.448,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021437.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508021458.496, "dur": 6.592,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021458.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508021465.216, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021465.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021466.560, "dur": 13.376,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021466.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508021480.032, "dur": 2.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021480.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021483.040, "dur": 16.832,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021483.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508021500.000, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021500.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508021501.376, "dur": 8.800,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021501.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508021510.304, "dur": 3.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021510.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508021513.504, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021513.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508021514.688, "dur": 15.232,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021514.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508021530.048, "dur": 1.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021530.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021531.200, "dur": 13.792,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021531.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508021545.088, "dur": 2.272,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021545.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021547.456, "dur": 3.648,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021547.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508021551.264, "dur": 0.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021551.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508021552.352, "dur": 152.128,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021552.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508021704.608, "dur": 7.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021704.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508021711.840, "dur": 4.736,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021711.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508021716.736, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021716.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508021718.368, "dur": 12.032,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021718.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508021730.528, "dur": 6.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021730.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508021737.568, "dur": 1.728,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021737.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508021739.392, "dur": 7.648,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021739.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508021747.136, "dur": 8.864,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021747.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508021756.128, "dur": 5.824,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021756.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021762.080, "dur": 22.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021762.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508021784.960, "dur": 3.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021784.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021788.224, "dur": 27.008,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021788.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508021815.360, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021815.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508021817.024, "dur": 25.888,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021817.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508021843.040, "dur": 6.752,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021843.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508021849.920, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021849.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021851.296, "dur": 23.264,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021851.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508021874.688, "dur": 3.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021874.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021877.888, "dur": 30.176,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021877.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508021908.192, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021908.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508021909.888, "dur": 31.488,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021909.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508021943.328, "dur": 1.024,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021943.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508021949.184, "dur": 52.544,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508021949.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022001.856, "dur": 48.512,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022001.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508022050.496, "dur": 2.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022050.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508022052.704, "dur": 21.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022052.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508022074.336, "dur": 6.720,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022074.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508022081.152, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022081.152,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022082.528, "dur": 13.088,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022082.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508022095.744, "dur": 2.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022095.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022098.720, "dur": 16.416,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022098.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508022115.264, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022115.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508022116.608, "dur": 9.152,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022116.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508022125.888, "dur": 3.232,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022125.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508022129.248, "dur": 1.089,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022129.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508022130.464, "dur": 14.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022130.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508022145.408, "dur": 1.025,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022145.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022146.560, "dur": 13.793,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022146.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508022160.513, "dur": 2.239,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022160.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022162.880, "dur": 3.680,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022162.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508022166.656, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022166.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508022167.744, "dur": 155.872,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022167.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508022323.744, "dur": 7.200,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022323.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508022331.072, "dur": 4.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022331.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508022336.128, "dur": 1.505,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022336.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508022337.760, "dur": 11.617,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022337.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508022349.504, "dur": 7.264,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022349.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508022356.896, "dur": 1.792,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022356.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508022358.784, "dur": 7.840,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022358.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508022366.752, "dur": 9.601,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022366.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508022376.448, "dur": 6.177,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022376.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022382.752, "dur": 22.881,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022382.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508022405.760, "dur": 3.169,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022405.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022409.056, "dur": 26.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022409.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508022436.129, "dur": 1.503,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022436.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508022437.760, "dur": 24.480,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022437.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508022462.369, "dur": 6.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022462.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508022469.344, "dur": 1.249,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022469.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022470.688, "dur": 22.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022470.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508022493.696, "dur": 3.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022493.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022496.897, "dur": 29.567,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022496.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508022526.593, "dur": 1.631,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022526.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508022528.320, "dur": 32.001,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022528.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508022561.696, "dur": 1.056,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022561.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022567.713, "dur": 50.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022567.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022618.784, "dur": 48.352,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022618.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508022667.265, "dur": 2.016,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022667.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508022669.441, "dur": 21.792,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022669.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508022691.360, "dur": 6.625,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022691.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508022698.112, "dur": 1.249,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022698.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022699.488, "dur": 13.280,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022699.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508022712.897, "dur": 2.847,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022712.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022715.873, "dur": 16.416,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022715.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508022732.449, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022732.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508022733.793, "dur": 9.695,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022733.793,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508022743.648, "dur": 3.041,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022743.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508022746.816, "dur": 1.057,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022746.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508022747.969, "dur": 15.200,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022747.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508022763.265, "dur": 1.183,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022763.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022764.576, "dur": 13.409,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022764.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508022778.112, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022778.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022780.481, "dur": 3.583,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022780.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508022784.193, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022784.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508022785.280, "dur": 155.616,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022785.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508022941.056, "dur": 7.297,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022941.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508022948.480, "dur": 4.801,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022948.480,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508022953.376, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022953.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508022955.041, "dur": 11.680,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022955.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508022966.849, "dur": 7.199,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022966.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508022974.177, "dur": 1.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022974.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508022976.033, "dur": 7.712,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022976.033,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508022983.872, "dur": 9.184,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022983.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508022993.216, "dur": 6.017,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022993.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508022999.361, "dur": 22.720,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508022999.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023022.176, "dur": 3.169,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023022.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023025.472, "dur": 26.625,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023025.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508023052.192, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023052.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508023053.824, "dur": 20.641,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023053.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508023074.560, "dur": 7.073,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023074.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023081.728, "dur": 1.249,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023081.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508023083.073, "dur": 18.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023083.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508023102.081, "dur": 13.440,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023102.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023115.649, "dur": 1.247,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023115.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023117.025, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023117.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508023118.369, "dur": 25.663,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023118.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023144.192, "dur": 1.217,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023144.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023145.537, "dur": 1.215,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023145.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508023146.881, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023146.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508023148.065, "dur": 36.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023148.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508023185.152, "dur": 9.505,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023185.152,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023194.785, "dur": 1.824,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023194.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023196.736, "dur": 10.561,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023196.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023207.425, "dur": 2.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023207.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023210.496, "dur": 11.777,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023210.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508023222.433, "dur": 1.152,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023222.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508023223.681, "dur": 7.583,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023223.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023231.361, "dur": 1.280,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023231.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508023232.737, "dur": 4.416,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023232.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023237.281, "dur": 1.215,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023237.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023238.625, "dur": 10.272,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023238.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023249.025, "dur": 2.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023249.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023252.001, "dur": 10.464,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023252.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508023262.560, "dur": 1.185,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023262.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290508023263.873, "dur": 2.528,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023263.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290508023266.529, "dur": 7.424,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023266.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023274.049, "dur": 22.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023274.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023297.088, "dur": 3.009,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023297.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023300.257, "dur": 29.760,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023300.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508023330.145, "dur": 1.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023330.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508023331.873, "dur": 31.712,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023331.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508023365.601, "dur": 1.087,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.13247470101195952
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023365.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023371.553, "dur": 53.664,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023371.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023425.345, "dur": 48.576,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023425.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508023474.081, "dur": 2.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023474.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508023476.289, "dur": 29.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023476.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508023505.889, "dur": 6.752,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023505.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023512.769, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023512.769,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023514.145, "dur": 13.184,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023514.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023527.425, "dur": 2.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023527.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023530.465, "dur": 16.096,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023530.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508023546.689, "dur": 1.280,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023546.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508023548.097, "dur": 9.312,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023548.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508023557.537, "dur": 2.976,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023557.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508023560.673, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023560.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508023561.857, "dur": 15.360,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023561.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508023577.345, "dur": 1.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023577.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023578.497, "dur": 13.760,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023578.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023592.385, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023592.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023594.753, "dur": 3.680,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023594.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508023598.529, "dur": 0.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023598.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508023599.617, "dur": 162.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023599.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508023761.857, "dur": 6.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023761.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508023768.897, "dur": 4.800,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023768.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508023773.825, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023773.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508023775.457, "dur": 12.800,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023775.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508023788.353, "dur": 6.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023788.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508023795.297, "dur": 1.792,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023795.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508023797.217, "dur": 7.584,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023797.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508023804.929, "dur": 9.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023804.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508023814.593, "dur": 6.112,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023814.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023820.833, "dur": 23.296,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023820.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023844.257, "dur": 3.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023844.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023847.457, "dur": 27.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023847.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508023874.753, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023874.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508023876.385, "dur": 23.680,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023876.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508023900.193, "dur": 6.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023900.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508023907.169, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023907.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023908.545, "dur": 22.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023908.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508023931.617, "dur": 3.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023931.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508023934.785, "dur": 29.856,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023934.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508023964.801, "dur": 1.664,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023964.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508023966.625, "dur": 31.968,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508023966.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508024000.545, "dur": 1.024,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024000.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024006.401, "dur": 50.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024006.401,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024057.313, "dur": 48.352,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024057.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508024105.793, "dur": 2.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024105.793,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508024107.969, "dur": 20.608,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024107.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508024128.705, "dur": 6.592,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024128.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508024135.393, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024135.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024136.737, "dur": 12.896,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024136.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508024149.729, "dur": 2.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024149.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024152.705, "dur": 17.312,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024152.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508024170.145, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024170.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508024171.489, "dur": 9.440,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024171.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508024181.057, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024181.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508024184.289, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024184.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508024185.473, "dur": 15.872,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024185.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508024201.473, "dur": 1.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024201.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024203.137, "dur": 13.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024203.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508024216.801, "dur": 2.208,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024216.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024219.105, "dur": 3.584,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024219.105,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508024222.785, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024222.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508024223.873, "dur": 165.088,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024223.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508024389.057, "dur": 6.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024389.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508024396.097, "dur": 4.768,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024396.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508024400.993, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024400.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508024402.625, "dur": 11.424,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024402.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508024414.177, "dur": 6.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024414.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508024421.153, "dur": 1.664,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024421.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508024422.945, "dur": 7.744,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024422.945,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508024430.849, "dur": 9.760,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024430.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508024440.737, "dur": 6.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024440.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024446.977, "dur": 22.400,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024446.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508024469.505, "dur": 3.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024469.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024472.801, "dur": 27.008,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024472.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508024499.905, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024499.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508024501.569, "dur": 25.376,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024501.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508024527.041, "dur": 6.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024527.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508024534.081, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024534.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024535.457, "dur": 23.456,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024535.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508024559.009, "dur": 3.200,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024559.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024562.305, "dur": 30.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024562.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508024592.481, "dur": 1.632,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024592.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508024594.209, "dur": 31.872,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024594.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508024629.313, "dur": 1.056,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024629.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024635.105, "dur": 52.768,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024635.105,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024688.001, "dur": 48.640,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024688.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508024736.801, "dur": 2.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024736.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508024739.009, "dur": 22.112,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024739.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508024761.217, "dur": 6.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024761.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508024768.129, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024768.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024769.505, "dur": 13.376,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024769.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508024783.009, "dur": 2.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024783.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024785.953, "dur": 16.096,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024785.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508024802.177, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024802.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508024803.553, "dur": 9.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024803.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508024812.897, "dur": 3.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024812.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508024816.193, "dur": 1.088,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024816.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508024817.409, "dur": 15.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024817.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508024832.641, "dur": 1.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024832.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024833.793, "dur": 13.792,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024833.793,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508024847.713, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024847.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508024850.081, "dur": 3.648,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024850.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508024853.857, "dur": 0.992,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024853.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508024854.977, "dur": 155.968,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508024854.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508025011.073, "dur": 7.008,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025011.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508025018.209, "dur": 4.832,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025018.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508025023.137, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025023.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508025024.769, "dur": 11.584,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025024.769,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508025036.513, "dur": 6.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025036.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508025043.489, "dur": 1.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025043.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508025045.345, "dur": 8.032,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025045.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508025053.505, "dur": 10.208,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025053.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508025063.841, "dur": 6.016,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025063.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025070.017, "dur": 22.112,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025070.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508025092.257, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025092.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025095.489, "dur": 26.752,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025095.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508025122.369, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025122.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508025124.033, "dur": 23.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025124.033,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508025147.297, "dur": 6.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025147.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508025154.177, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025154.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025155.553, "dur": 23.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025155.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508025178.753, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025178.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025181.985, "dur": 29.824,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025181.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508025211.937, "dur": 1.632,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025211.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508025213.665, "dur": 31.808,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025213.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508025247.713, "dur": 0.992,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025247.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025253.537, "dur": 51.520,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025253.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025305.185, "dur": 47.936,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025305.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508025353.249, "dur": 2.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025353.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508025355.425, "dur": 21.184,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025355.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508025376.737, "dur": 6.528,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025376.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508025383.393, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025383.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025384.737, "dur": 13.152,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025384.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508025398.017, "dur": 2.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025398.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025400.993, "dur": 16.544,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025400.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508025417.633, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025417.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508025419.009, "dur": 8.672,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025419.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508025427.809, "dur": 3.232,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025427.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508025431.169, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025431.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508025432.353, "dur": 14.752,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025432.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508025447.201, "dur": 1.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025447.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025448.353, "dur": 13.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025448.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508025462.049, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025462.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025464.385, "dur": 3.616,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025464.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508025468.161, "dur": 0.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025468.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508025469.217, "dur": 159.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025469.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508025628.449, "dur": 6.944,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025628.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508025635.521, "dur": 4.896,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025635.521,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508025640.545, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025640.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508025642.177, "dur": 12.608,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025642.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508025654.913, "dur": 6.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025654.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508025661.793, "dur": 1.728,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025661.793,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508025663.617, "dur": 7.520,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025663.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508025671.265, "dur": 8.288,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025671.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508025679.681, "dur": 5.984,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025679.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025685.793, "dur": 22.304,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025685.793,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508025708.257, "dur": 3.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025708.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025711.553, "dur": 27.392,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025711.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508025739.073, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025739.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508025740.705, "dur": 23.456,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025740.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508025764.289, "dur": 6.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025764.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508025771.201, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025771.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025772.577, "dur": 22.688,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025772.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508025795.393, "dur": 3.200,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025795.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025798.689, "dur": 29.952,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025798.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508025828.737, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025828.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508025830.465, "dur": 31.424,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025830.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508025864.321, "dur": 1.152,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025864.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025870.241, "dur": 54.272,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025870.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508025924.609, "dur": 48.608,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025924.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508025973.345, "dur": 2.112,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025973.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508025975.585, "dur": 21.376,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025975.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508025997.089, "dur": 6.688,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508025997.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508026003.905, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026003.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026005.249, "dur": 13.312,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026005.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508026018.657, "dur": 2.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026018.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026021.665, "dur": 16.384,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026021.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508026038.177, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026038.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508026039.553, "dur": 9.728,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026039.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508026049.409, "dur": 3.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026049.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508026052.705, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026052.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508026053.889, "dur": 15.808,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026053.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508026069.825, "dur": 0.992,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026069.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026070.977, "dur": 13.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026070.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508026084.705, "dur": 2.272,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026084.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026087.105, "dur": 3.648,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026087.105,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508026090.881, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026090.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508026091.969, "dur": 159.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026091.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508026251.169, "dur": 7.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026251.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508026258.465, "dur": 4.832,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026258.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508026263.393, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026263.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508026265.025, "dur": 11.968,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026265.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508026277.089, "dur": 6.752,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026277.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508026283.969, "dur": 1.728,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026283.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508026285.825, "dur": 8.736,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026285.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508026294.689, "dur": 10.560,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026294.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508026305.377, "dur": 5.984,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026305.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026311.489, "dur": 22.560,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026311.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508026334.177, "dur": 3.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026334.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026337.377, "dur": 27.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026337.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508026364.577, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026364.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508026366.209, "dur": 22.016,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026366.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508026388.353, "dur": 6.784,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026388.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508026395.265, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026395.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026396.609, "dur": 22.912,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026396.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508026419.649, "dur": 3.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026419.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026422.913, "dur": 29.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026422.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508026452.641, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026452.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508026454.337, "dur": 31.392,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026454.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508026487.873, "dur": 0.992,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026487.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026493.729, "dur": 52.448,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026493.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026546.305, "dur": 47.808,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026546.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508026594.241, "dur": 2.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026594.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508026596.417, "dur": 21.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026596.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508026617.601, "dur": 6.592,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026617.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508026624.321, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026624.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026625.665, "dur": 13.120,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026625.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508026638.913, "dur": 2.848,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026638.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026641.889, "dur": 16.512,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026641.889,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508026658.529, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026658.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508026659.873, "dur": 9.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026659.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508026669.057, "dur": 3.104,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026669.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508026672.289, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026672.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508026673.473, "dur": 15.712,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026673.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508026689.281, "dur": 1.024,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026689.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026690.433, "dur": 13.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026690.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508026704.065, "dur": 2.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026704.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026706.433, "dur": 3.584,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026706.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508026710.145, "dur": 0.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026710.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508026711.201, "dur": 160.608,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026711.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508026871.937, "dur": 7.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026871.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508026879.169, "dur": 4.896,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026879.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508026884.161, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026884.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508026885.794, "dur": 11.935,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026885.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508026897.857, "dur": 6.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026897.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508026904.801, "dur": 1.664,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026904.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508026906.625, "dur": 7.552,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026906.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508026914.305, "dur": 9.984,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026914.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508026924.417, "dur": 6.176,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026924.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026930.721, "dur": 22.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026930.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508026953.665, "dur": 3.168,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026953.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508026956.961, "dur": 27.200,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026956.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508026984.321, "dur": 1.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026984.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508026985.953, "dur": 24.192,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508026985.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508027010.273, "dur": 6.688,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027010.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027017.057, "dur": 1.280,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027017.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027018.433, "dur": 23.040,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027018.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508027041.569, "dur": 3.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027041.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027044.801, "dur": 30.048,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027044.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508027074.977, "dur": 1.600,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027074.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508027076.705, "dur": 31.488,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027076.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508027110.497, "dur": 1.025,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.1404878048780488
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027110.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027116.257, "dur": 50.240,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027116.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027166.625, "dur": 48.385,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027166.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508027215.137, "dur": 2.049,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027215.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508027217.313, "dur": 20.896,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027217.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508027238.337, "dur": 6.656,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027238.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027245.121, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027245.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027246.497, "dur": 13.249,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027246.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508027259.842, "dur": 2.847,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027259.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027262.818, "dur": 17.567,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027262.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508027280.514, "dur": 1.247,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027280.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508027281.857, "dur": 9.344,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027281.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508027291.329, "dur": 2.880,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027291.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508027294.337, "dur": 1.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027294.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508027295.521, "dur": 14.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027295.521,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508027310.433, "dur": 1.664,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027310.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027312.226, "dur": 13.760,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027312.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508027326.146, "dur": 2.207,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027326.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027328.482, "dur": 3.583,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027328.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508027332.193, "dur": 0.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027332.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508027333.249, "dur": 161.952,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027333.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508027495.394, "dur": 6.943,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027495.394,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508027502.466, "dur": 4.832,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027502.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508027507.458, "dur": 1.503,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027507.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508027509.058, "dur": 11.647,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027509.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508027520.802, "dur": 7.071,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027520.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508027527.970, "dur": 1.696,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027527.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508027529.761, "dur": 7.360,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027529.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508027537.249, "dur": 8.928,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027537.249,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508027546.306, "dur": 5.888,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027546.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027552.321, "dur": 22.528,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027552.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508027574.978, "dur": 3.167,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027574.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027578.241, "dur": 26.720,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027578.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508027605.089, "dur": 1.505,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027605.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508027606.721, "dur": 23.072,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027606.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508027629.922, "dur": 6.911,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027629.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027636.961, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027636.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508027638.306, "dur": 19.039,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027638.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508027657.442, "dur": 13.440,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027657.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027671.009, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027671.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027672.354, "dur": 1.247,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027672.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508027673.729, "dur": 26.049,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027673.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027699.905, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027699.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027701.218, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027701.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508027702.561, "dur": 1.089,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027702.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508027703.810, "dur": 36.831,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027703.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508027740.769, "dur": 13.569,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027740.769,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027754.241, "dur": 1.217,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027754.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027755.585, "dur": 10.817,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027755.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508027766.529, "dur": 2.945,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027766.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027769.601, "dur": 11.648,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027769.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508027781.345, "dur": 1.185,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027781.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508027782.625, "dur": 7.488,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027782.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027790.209, "dur": 1.249,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027790.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508027791.586, "dur": 4.575,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027791.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508027796.258, "dur": 1.216,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027796.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027797.570, "dur": 10.175,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027797.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508027807.874, "dur": 2.816,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027807.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027810.850, "dur": 10.368,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027810.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508027821.378, "dur": 1.151,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027821.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290508027822.658, "dur": 2.527,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027822.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290508027825.314, "dur": 8.095,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027825.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027833.538, "dur": 22.879,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027833.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508027856.546, "dur": 3.136,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027856.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027859.809, "dur": 29.985,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027859.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508027889.922, "dur": 1.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027889.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508027891.618, "dur": 31.519,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027891.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508027924.546, "dur": 1.087,
    "args": {
      "External id": 2842,      "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "bytes": 144, "memory bandwidth (GB/s)": 0.13247470101195952
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027924.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027930.434, "dur": 53.504,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027930.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508027984.065, "dur": 48.577,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508027984.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508028032.770, "dur": 2.112,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028032.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290508028035.009, "dur": 33.249,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028035.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028068.354, "dur": 16.896,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028068.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508028085.377, "dur": 9.121,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028085.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508028094.625, "dur": 17.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028094.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028111.810, "dur": 5.056,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028111.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508028116.993, "dur": 157.825,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028116.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508028274.978, "dur": 13.440,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028274.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508028288.514, "dur": 7.488,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028288.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508028296.129, "dur": 6.241,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028296.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508028302.497, "dur": 4.896,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028302.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508028307.522, "dur": 5.728,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028307.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028313.378, "dur": 27.488,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028313.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290508028340.994, "dur": 19.999,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028340.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508028361.122, "dur": 1.248,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028361.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028362.498, "dur": 13.856,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028362.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508028376.449, "dur": 2.913,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028376.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508028379.490, "dur": 1.824,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028379.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508028381.442, "dur": 3.584,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028381.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508028385.154, "dur": 1.088,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028385.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508028386.370, "dur": 1.376,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028386.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028387.906, "dur": 13.472,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028387.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508028401.538, "dur": 2.080,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028401.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508028403.746, "dur": 0.960,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028403.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508028404.834, "dur": 4.832,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028404.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508028409.794, "dur": 1.568,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028409.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508028411.458, "dur": 7.360,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028411.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508028418.946, "dur": 1.728,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028418.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508028420.770, "dur": 22.208,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028420.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508028443.073, "dur": 3.041,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028443.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508028446.242, "dur": 1.536,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028446.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290508028447.905, "dur": 25.121,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028447.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508028473.154, "dur": 1.440,
    "args": {
      "External id": 2842,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 505257,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 505257, "pid": 0, "tid": 7, "ts": 1290508028473.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507984846.759, "dur": 919.894,
    "args": {
      "External id": 2842,      "cbid": 311, "correlation": 505257
    }
  },
  {
    "ph": "s", "id": 505257, "pid": 14565, "tid": 14640, "ts": 1290507984846.759,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507989798.787, "dur": 1.033,
    "args": {
      "External id": 3610,      "cbid": 317, "correlation": 506032
    }
  },
  {
    "ph": "f", "id": 506032, "pid": 14565, "tid": 14640, "ts": 1290507989798.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290508028477.921, "dur": 8.608,
    "args": {
      "External id": 3610,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506034,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506034, "pid": 0, "tid": 7, "ts": 1290508028477.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290508028486.882, "dur": 15.968,
    "args": {
      "External id": 3610,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506034,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506034, "pid": 0, "tid": 7, "ts": 1290508028486.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507989800.511, "dur": 15.716,
    "args": {
      "External id": 3610,      "cbid": 311, "correlation": 506034
    }
  },
  {
    "ph": "s", "id": 506034, "pid": 14565, "tid": 14640, "ts": 1290507989800.511,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290508028505.858, "dur": 22.368,
    "args": {
      "External id": 3612,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506045,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506045, "pid": 0, "tid": 7, "ts": 1290508028505.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507989864.987, "dur": 11.387,
    "args": {
      "External id": 3612,      "cbid": 211, "correlation": 506045
    }
  },
  {
    "ph": "s", "id": 506045, "pid": 14565, "tid": 14640, "ts": 1290507989864.987,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290508034948.867, "dur": 2.080,
    "args": {
      "External id": 3620,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506130,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506130, "pid": 0, "tid": 7, "ts": 1290508034948.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507993625.157, "dur": 19.831,
    "args": {
      "External id": 3620,      "cbid": 211, "correlation": 506130
    }
  },
  {
    "ph": "s", "id": 506130, "pid": 14565, "tid": 14640, "ts": 1290507993625.157,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507993654.270, "dur": 0.807,
    "args": {
      "External id": 3618,      "cbid": 317, "correlation": 506138
    }
  },
  {
    "ph": "f", "id": 506138, "pid": 14565, "tid": 14640, "ts": 1290507993654.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290508034955.075, "dur": 12.128,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508034955.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290508034967.331, "dur": 18.048,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508034967.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290508034985.475, "dur": 6.528,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508034985.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508034992.099, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508034992.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508034993.475, "dur": 23.488,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508034993.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508035017.091, "dur": 3.264,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035017.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035020.451, "dur": 30.400,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035020.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508035050.979, "dur": 1.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035050.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508035052.739, "dur": 33.152,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035052.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508035088.419, "dur": 1.088,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035088.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035094.307, "dur": 56.736,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035094.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035151.139, "dur": 48.576,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035151.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508035199.843, "dur": 2.048,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035199.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290508035202.019, "dur": 29.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035202.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508035231.715, "dur": 6.496,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035231.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508035238.339, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035238.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035239.715, "dur": 13.344,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035239.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508035253.187, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035253.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035256.195, "dur": 16.160,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035256.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508035272.515, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035272.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508035273.859, "dur": 9.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035273.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508035283.267, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035283.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508035286.499, "dur": 1.088,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035286.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508035287.715, "dur": 15.232,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035287.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508035303.075, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035303.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035304.227, "dur": 13.824,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035304.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508035318.179, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035318.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035320.547, "dur": 3.712,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035320.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508035324.387, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035324.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508035325.475, "dur": 159.808,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035325.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508035485.411, "dur": 7.008,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035485.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508035492.547, "dur": 4.800,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035492.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508035497.475, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035497.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508035499.107, "dur": 11.712,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035499.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508035510.947, "dur": 7.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035510.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508035518.179, "dur": 1.664,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035518.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508035519.971, "dur": 7.552,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035519.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508035527.651, "dur": 8.704,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035527.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508035536.515, "dur": 5.888,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035536.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035542.499, "dur": 23.328,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035542.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508035565.923, "dur": 3.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035565.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035569.187, "dur": 26.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035569.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508035596.195, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035596.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508035597.827, "dur": 19.968,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035597.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508035617.923, "dur": 6.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035617.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508035624.899, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035624.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035626.275, "dur": 22.272,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035626.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508035648.643, "dur": 3.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035648.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035651.811, "dur": 29.760,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035651.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508035681.699, "dur": 1.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035681.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508035683.395, "dur": 32.128,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035683.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508035717.091, "dur": 1.024,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035717.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035723.043, "dur": 49.984,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035723.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035773.155, "dur": 48.320,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035773.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508035821.635, "dur": 2.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035821.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508035823.907, "dur": 21.760,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035823.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508035845.795, "dur": 6.624,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035845.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508035852.547, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035852.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035853.923, "dur": 13.408,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035853.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508035867.459, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035867.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035870.499, "dur": 16.704,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035870.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508035887.363, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035887.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508035888.739, "dur": 9.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035888.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508035898.403, "dur": 3.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035898.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508035901.603, "dur": 1.088,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035901.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508035902.787, "dur": 14.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035902.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508035917.763, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035917.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035918.915, "dur": 13.920,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035918.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508035932.931, "dur": 2.304,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035932.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508035935.331, "dur": 3.648,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035935.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508035939.139, "dur": 0.928,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035939.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508035940.195, "dur": 158.592,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508035940.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508036098.947, "dur": 7.296,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036098.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508036106.339, "dur": 4.864,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036106.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508036111.331, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036111.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508036112.963, "dur": 12.128,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036112.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508036125.219, "dur": 6.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036125.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508036132.323, "dur": 1.664,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036132.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508036134.115, "dur": 7.328,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036134.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508036141.571, "dur": 9.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036141.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508036151.395, "dur": 5.888,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036151.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036157.411, "dur": 22.048,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036157.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508036179.555, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036179.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036182.787, "dur": 27.008,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036182.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508036209.923, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036209.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508036211.555, "dur": 25.088,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036211.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508036236.739, "dur": 6.720,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036236.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508036243.587, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036243.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036244.963, "dur": 22.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036244.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508036267.939, "dur": 3.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036267.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036271.139, "dur": 29.984,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036271.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508036301.219, "dur": 1.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036301.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508036302.915, "dur": 31.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036302.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508036335.555, "dur": 1.120,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036335.555,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036341.475, "dur": 50.272,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036341.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036391.875, "dur": 48.224,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036391.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508036440.259, "dur": 2.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036440.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508036442.499, "dur": 21.408,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036442.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508036464.035, "dur": 6.592,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036464.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508036470.755, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036470.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036472.099, "dur": 13.312,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036472.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508036485.539, "dur": 2.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036485.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036488.515, "dur": 16.608,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036488.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508036505.251, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036505.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508036506.595, "dur": 9.472,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036506.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508036516.195, "dur": 2.688,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036516.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508036519.043, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036519.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508036520.227, "dur": 15.680,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036520.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508036536.035, "dur": 1.184,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036536.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036537.411, "dur": 13.376,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036537.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508036550.915, "dur": 2.208,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036550.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036553.251, "dur": 3.680,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036553.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508036557.059, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036557.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508036558.115, "dur": 158.464,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036558.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508036716.739, "dur": 6.944,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036716.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508036723.811, "dur": 5.120,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036723.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508036729.059, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036729.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508036730.691, "dur": 11.904,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036730.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508036742.723, "dur": 6.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036742.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508036749.827, "dur": 1.664,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036749.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508036751.619, "dur": 7.616,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036751.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508036759.363, "dur": 8.192,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036759.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508036767.683, "dur": 5.856,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036767.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036773.635, "dur": 22.496,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036773.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508036796.259, "dur": 3.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036796.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036799.523, "dur": 27.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036799.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508036826.723, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036826.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508036828.387, "dur": 25.376,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036828.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508036853.923, "dur": 6.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036853.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508036860.931, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036860.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036862.275, "dur": 22.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036862.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508036885.379, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036885.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036888.611, "dur": 29.952,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036888.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508036918.691, "dur": 1.664,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036918.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508036920.451, "dur": 31.872,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036920.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508036954.051, "dur": 1.024,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036954.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508036959.843, "dur": 52.640,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508036959.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037012.579, "dur": 48.513,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037012.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508037061.219, "dur": 2.049,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037061.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508037063.395, "dur": 21.632,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037063.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508037085.156, "dur": 6.591,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037085.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508037091.843, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037091.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037093.219, "dur": 13.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037093.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508037106.627, "dur": 2.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037106.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037109.603, "dur": 16.193,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037109.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508037125.923, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037125.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508037127.331, "dur": 9.312,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037127.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508037136.803, "dur": 3.008,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037136.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508037139.971, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037139.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508037141.155, "dur": 14.688,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037141.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508037155.939, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037155.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037157.091, "dur": 13.792,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037157.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508037171.043, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037171.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037173.411, "dur": 3.649,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037173.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508037177.187, "dur": 0.928,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037177.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508037178.243, "dur": 160.864,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037178.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508037339.267, "dur": 6.945,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037339.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508037346.339, "dur": 4.832,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037346.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508037351.267, "dur": 1.537,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037351.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508037352.931, "dur": 11.488,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037352.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508037364.547, "dur": 6.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037364.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508037371.652, "dur": 1.695,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037371.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508037373.475, "dur": 7.808,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037373.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508037381.380, "dur": 9.984,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037381.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508037391.459, "dur": 6.017,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037391.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037397.571, "dur": 22.720,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037397.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508037420.451, "dur": 3.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037420.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037423.651, "dur": 26.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037423.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508037450.691, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037450.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508037452.356, "dur": 23.135,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037452.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508037475.587, "dur": 6.752,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037475.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508037482.435, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037482.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037483.812, "dur": 23.135,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037483.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508037507.076, "dur": 3.103,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037507.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037510.308, "dur": 29.376,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037510.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508037539.811, "dur": 1.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037539.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508037541.539, "dur": 31.777,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037541.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508037574.627, "dur": 1.056,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037574.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037580.515, "dur": 50.241,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037580.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037630.883, "dur": 48.417,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037630.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508037679.427, "dur": 2.049,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037679.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508037681.603, "dur": 20.865,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037681.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508037702.595, "dur": 6.592,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037702.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508037709.316, "dur": 1.247,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037709.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037710.660, "dur": 13.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037710.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508037723.843, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037723.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037726.852, "dur": 16.832,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037726.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508037743.811, "dur": 1.217,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037743.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508037745.155, "dur": 9.377,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037745.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508037754.692, "dur": 2.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037754.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508037757.795, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037757.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508037758.980, "dur": 14.784,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037758.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508037773.891, "dur": 1.153,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037773.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037775.140, "dur": 13.471,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037775.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508037788.707, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037788.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508037791.076, "dur": 3.647,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037791.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508037794.852, "dur": 0.927,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037794.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508037795.908, "dur": 166.752,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037795.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508037962.788, "dur": 7.008,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037962.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508037969.923, "dur": 4.832,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037969.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508037974.915, "dur": 1.505,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037974.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508037976.548, "dur": 11.520,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037976.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508037988.195, "dur": 7.137,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037988.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508037995.460, "dur": 1.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037995.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508037997.283, "dur": 7.488,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508037997.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508038004.900, "dur": 9.344,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038004.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508038014.339, "dur": 6.017,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038014.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038020.451, "dur": 22.944,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038020.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508038043.492, "dur": 3.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038043.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038046.787, "dur": 26.977,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038046.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508038073.860, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038073.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508038075.491, "dur": 28.545,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038075.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508038104.132, "dur": 6.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038104.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508038111.140, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038111.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038112.515, "dur": 23.137,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038112.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508038135.812, "dur": 3.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038135.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038138.980, "dur": 30.016,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038138.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508038169.124, "dur": 1.632,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038169.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508038170.884, "dur": 31.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038170.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508038203.459, "dur": 1.120,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038203.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038209.444, "dur": 51.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038209.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038260.612, "dur": 48.576,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038260.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508038309.316, "dur": 2.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038309.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508038311.556, "dur": 21.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038311.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508038332.708, "dur": 6.591,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038332.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508038339.428, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038339.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038340.804, "dur": 13.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038340.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508038354.179, "dur": 2.849,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038354.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038357.155, "dur": 16.769,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038357.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508038374.052, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038374.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508038375.427, "dur": 8.929,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038375.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508038384.484, "dur": 2.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038384.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508038387.588, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038387.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508038388.772, "dur": 14.528,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038388.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508038403.396, "dur": 1.184,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038403.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038404.676, "dur": 13.856,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038404.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508038418.628, "dur": 2.271,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038418.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038421.028, "dur": 3.712,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038421.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508038424.868, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038424.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508038425.956, "dur": 152.544,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038425.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508038578.628, "dur": 7.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038578.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508038585.796, "dur": 4.896,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038585.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508038590.820, "dur": 1.472,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038590.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508038592.452, "dur": 11.840,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038592.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508038604.452, "dur": 6.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038604.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508038611.556, "dur": 1.728,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038611.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508038613.380, "dur": 7.520,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038613.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508038620.996, "dur": 8.384,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038620.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508038629.508, "dur": 5.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038629.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038635.204, "dur": 22.016,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038635.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508038657.316, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038657.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038660.580, "dur": 26.528,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038660.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508038687.236, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038687.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508038688.900, "dur": 25.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038688.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508038714.628, "dur": 6.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038714.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508038721.668, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038721.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038723.012, "dur": 23.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038723.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508038746.340, "dur": 3.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038746.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038749.540, "dur": 29.856,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038749.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508038779.556, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038779.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508038781.220, "dur": 31.648,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038781.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508038814.756, "dur": 0.992,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038814.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038820.612, "dur": 50.592,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038820.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038871.332, "dur": 48.064,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038871.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508038919.524, "dur": 2.080,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038919.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508038921.732, "dur": 21.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038921.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508038943.524, "dur": 6.688,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038943.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508038950.340, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038950.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038951.684, "dur": 13.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038951.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508038965.060, "dur": 2.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038965.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508038968.004, "dur": 16.224,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038968.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508038984.356, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038984.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508038985.732, "dur": 9.344,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038985.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508038995.204, "dur": 3.424,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038995.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508038998.788, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038998.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508038999.972, "dur": 14.944,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508038999.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508039015.044, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039015.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039016.196, "dur": 13.440,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039016.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508039029.796, "dur": 2.208,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039029.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039032.100, "dur": 3.648,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039032.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508039035.876, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039035.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508039036.932, "dur": 154.144,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039036.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508039191.268, "dur": 7.008,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039191.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508039198.436, "dur": 4.928,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039198.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508039203.460, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039203.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508039205.124, "dur": 11.680,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039205.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508039216.932, "dur": 7.264,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039216.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508039224.324, "dur": 1.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039224.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508039226.148, "dur": 7.680,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039226.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508039233.956, "dur": 9.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039233.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508039243.652, "dur": 6.304,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039243.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039250.052, "dur": 22.688,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039250.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508039272.836, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039272.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039276.100, "dur": 26.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039276.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508039303.204, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039303.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508039304.836, "dur": 24.512,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039304.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508039329.508, "dur": 6.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039329.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508039336.516, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039336.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039337.892, "dur": 23.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039337.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508039361.188, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039361.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039364.420, "dur": 29.856,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039364.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508039394.404, "dur": 1.632,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039394.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508039396.164, "dur": 31.808,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039396.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508039431.268, "dur": 1.024,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039431.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039437.284, "dur": 50.528,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039437.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039487.940, "dur": 48.416,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039487.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508039536.484, "dur": 2.048,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039536.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508039538.660, "dur": 21.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039538.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508039560.036, "dur": 6.624,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039560.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508039566.756, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039566.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039568.132, "dur": 13.376,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039568.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508039581.636, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039581.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039584.644, "dur": 16.608,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039584.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508039601.380, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039601.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508039602.724, "dur": 9.120,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039602.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508039611.972, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039611.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508039614.980, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039614.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508039616.164, "dur": 14.656,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039616.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508039630.916, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039630.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039632.292, "dur": 14.016,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039632.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508039646.436, "dur": 2.208,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039646.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039648.772, "dur": 3.680,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039648.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508039652.580, "dur": 0.928,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039652.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508039653.636, "dur": 158.656,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039653.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508039812.580, "dur": 7.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039812.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508039819.812, "dur": 4.832,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039819.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508039824.740, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039824.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508039826.372, "dur": 11.808,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039826.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508039838.308, "dur": 6.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039838.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508039845.412, "dur": 1.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039845.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508039847.268, "dur": 7.744,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039847.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508039855.140, "dur": 9.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039855.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508039864.324, "dur": 6.080,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039864.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039870.564, "dur": 23.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039870.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508039893.764, "dur": 3.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039893.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508039897.028, "dur": 27.296,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039897.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508039924.420, "dur": 1.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039924.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508039926.084, "dur": 20.800,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039926.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508039946.980, "dur": 7.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039946.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508039954.148, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039954.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508039955.524, "dur": 18.688,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039955.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508039974.308, "dur": 13.440,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039974.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508039987.876, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039987.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508039989.220, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039989.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508039990.596, "dur": 25.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508039990.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508040016.292, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040016.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508040017.636, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040017.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508040019.012, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040019.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508040020.196, "dur": 37.664,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040020.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508040058.020, "dur": 10.272,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040058.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508040068.420, "dur": 1.760,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040068.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040070.308, "dur": 10.560,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040070.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508040080.996, "dur": 2.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040080.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040084.036, "dur": 12.384,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040084.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508040096.548, "dur": 1.152,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040096.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508040097.828, "dur": 7.648,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040097.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508040105.572, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040105.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508040106.948, "dur": 4.320,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040106.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508040111.396, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040111.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040112.740, "dur": 10.176,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040112.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508040123.044, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040123.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040126.020, "dur": 10.624,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040126.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508040136.772, "dur": 1.184,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040136.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290508040138.052, "dur": 2.528,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040138.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290508040140.676, "dur": 7.392,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040140.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040148.196, "dur": 22.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040148.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508040171.268, "dur": 3.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040171.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040174.436, "dur": 30.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040174.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508040204.676, "dur": 1.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040204.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508040206.404, "dur": 32.256,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040206.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508040241.220, "dur": 0.992,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040241.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040247.076, "dur": 53.152,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040247.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040300.356, "dur": 48.640,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040300.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508040349.124, "dur": 2.080,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040349.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508040351.332, "dur": 30.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040351.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508040381.540, "dur": 6.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040381.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508040388.484, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040388.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040389.860, "dur": 13.344,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040389.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508040403.300, "dur": 2.944,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040403.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040406.340, "dur": 16.288,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040406.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508040422.756, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040422.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508040424.132, "dur": 9.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040424.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508040433.316, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040433.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508040436.580, "dur": 1.088,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040436.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508040437.764, "dur": 14.752,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040437.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508040452.644, "dur": 1.344,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040452.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040454.084, "dur": 13.920,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040454.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508040468.164, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040468.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040470.532, "dur": 3.648,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040470.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508040474.308, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040474.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508040475.460, "dur": 156.768,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040475.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508040632.420, "dur": 6.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040632.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508040639.428, "dur": 4.768,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040639.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508040644.324, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040644.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508040645.924, "dur": 12.608,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040645.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508040658.660, "dur": 6.784,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040658.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508040665.540, "dur": 1.792,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040665.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508040667.460, "dur": 7.616,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040667.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508040675.204, "dur": 9.184,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040675.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508040684.548, "dur": 5.888,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040684.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040690.564, "dur": 22.336,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040690.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508040713.028, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040713.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040716.292, "dur": 26.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040716.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508040743.364, "dur": 1.472,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040743.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508040744.964, "dur": 23.424,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040744.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508040768.516, "dur": 6.784,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040768.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508040775.396, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040775.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040776.740, "dur": 23.200,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040776.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508040800.068, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040800.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040803.332, "dur": 29.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040803.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508040833.156, "dur": 1.664,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040833.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508040834.948, "dur": 31.296,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040834.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508040867.908, "dur": 1.184,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.12162162162162163
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040867.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040874.084, "dur": 50.272,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040874.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508040924.484, "dur": 48.384,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040924.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508040972.964, "dur": 2.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040972.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508040975.204, "dur": 20.576,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040975.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508040995.908, "dur": 6.528,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508040995.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508041002.564, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041002.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041003.908, "dur": 13.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041003.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508041017.092, "dur": 2.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041017.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041020.068, "dur": 16.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041020.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508041037.156, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041037.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508041038.532, "dur": 10.080,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041038.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508041048.740, "dur": 2.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041048.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508041051.844, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041051.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508041053.028, "dur": 15.232,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041053.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508041068.388, "dur": 1.728,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041068.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041070.244, "dur": 13.632,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041070.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508041084.004, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041084.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041086.372, "dur": 3.584,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041086.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508041090.084, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041090.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508041091.140, "dur": 170.592,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041091.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508041261.860, "dur": 7.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041261.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508041269.028, "dur": 4.800,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041269.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508041273.956, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041273.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508041275.588, "dur": 11.456,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041275.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508041287.172, "dur": 6.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041287.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508041294.212, "dur": 1.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041294.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508041296.036, "dur": 7.904,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041296.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508041304.068, "dur": 9.152,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041304.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508041313.348, "dur": 5.984,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041313.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041319.492, "dur": 22.592,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041319.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508041342.212, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041342.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041345.444, "dur": 27.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041345.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508041372.708, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041372.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508041374.372, "dur": 26.144,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041374.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508041400.612, "dur": 7.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041400.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508041407.748, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041407.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041409.124, "dur": 23.296,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041409.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508041432.516, "dur": 3.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041432.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041435.780, "dur": 29.408,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041435.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508041465.316, "dur": 1.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041465.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508041467.012, "dur": 30.496,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041467.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508041500.644, "dur": 0.992,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041500.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041506.468, "dur": 50.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041506.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041557.508, "dur": 48.512,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041557.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508041606.148, "dur": 2.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041606.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508041608.388, "dur": 22.144,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041608.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508041630.628, "dur": 6.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041630.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508041637.604, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041637.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041638.980, "dur": 13.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041638.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508041652.612, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041652.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041655.588, "dur": 16.000,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041655.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508041671.716, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041671.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508041673.092, "dur": 9.152,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041673.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508041682.340, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041682.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508041685.604, "dur": 1.088,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041685.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508041686.788, "dur": 15.456,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041686.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508041702.372, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041702.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041703.524, "dur": 13.792,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041703.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508041717.444, "dur": 2.208,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041717.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041719.812, "dur": 3.616,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041719.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508041723.556, "dur": 0.992,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041723.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508041724.644, "dur": 157.920,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041724.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508041882.660, "dur": 7.008,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041882.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508041889.796, "dur": 4.832,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041889.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508041894.756, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041894.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508041896.388, "dur": 11.904,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041896.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508041908.420, "dur": 6.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041908.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508041915.556, "dur": 1.728,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041915.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508041917.412, "dur": 8.128,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041917.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508041925.668, "dur": 10.145,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041925.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508041935.940, "dur": 5.952,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041935.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041942.052, "dur": 22.144,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041942.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508041964.324, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041964.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508041967.524, "dur": 27.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041967.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508041994.788, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041994.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508041996.420, "dur": 22.752,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508041996.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508042019.268, "dur": 6.784,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042019.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508042026.148, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042026.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042027.524, "dur": 23.232,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042027.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508042050.884, "dur": 3.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042050.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042054.052, "dur": 29.952,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042054.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508042084.100, "dur": 1.633,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042084.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508042085.860, "dur": 32.225,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042085.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508042121.284, "dur": 1.088,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042121.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042127.236, "dur": 51.776,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042127.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042179.140, "dur": 48.416,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042179.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508042227.685, "dur": 2.047,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042227.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508042229.861, "dur": 21.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042229.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508042251.236, "dur": 6.528,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042251.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508042257.861, "dur": 1.247,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042257.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042259.237, "dur": 13.120,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042259.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508042272.452, "dur": 2.912,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042272.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042275.460, "dur": 16.704,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042275.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508042292.292, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042292.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508042293.668, "dur": 8.865,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042293.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508042302.660, "dur": 3.424,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042302.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508042306.244, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042306.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508042307.428, "dur": 14.497,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042307.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508042322.052, "dur": 0.993,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042322.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042323.205, "dur": 13.823,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042323.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508042337.188, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042337.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042339.557, "dur": 3.647,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042339.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508042343.333, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042343.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508042344.420, "dur": 156.001,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042344.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508042500.548, "dur": 7.041,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042500.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508042507.716, "dur": 4.896,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042507.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508042512.772, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042512.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508042514.405, "dur": 12.640,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042514.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508042527.141, "dur": 6.752,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042527.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508042533.988, "dur": 1.761,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042533.988,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508042535.844, "dur": 7.520,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042535.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508042543.461, "dur": 8.671,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042543.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508042552.260, "dur": 6.144,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042552.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042558.533, "dur": 22.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042558.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508042581.508, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042581.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042584.772, "dur": 27.200,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042584.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508042612.101, "dur": 1.503,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042612.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508042613.764, "dur": 23.617,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042613.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508042637.508, "dur": 6.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042637.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508042644.516, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042644.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042645.860, "dur": 23.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042645.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508042669.157, "dur": 3.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042669.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042672.356, "dur": 29.856,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042672.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508042702.341, "dur": 1.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042702.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508042704.068, "dur": 31.296,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042704.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508042736.613, "dur": 1.024,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042736.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042742.565, "dur": 50.688,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042742.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042793.380, "dur": 48.545,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042793.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508042842.052, "dur": 2.176,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042842.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508042844.357, "dur": 21.759,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042844.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508042866.245, "dur": 6.624,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042866.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508042872.997, "dur": 1.247,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042872.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042874.340, "dur": 13.377,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042874.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508042887.844, "dur": 2.849,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042887.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042890.820, "dur": 16.865,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042890.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508042907.812, "dur": 1.249,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042907.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508042909.157, "dur": 9.408,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042909.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508042918.692, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042918.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508042921.957, "dur": 1.055,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042921.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508042923.141, "dur": 15.327,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042923.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508042938.597, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042938.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042939.781, "dur": 13.887,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042939.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508042953.797, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042953.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508042956.164, "dur": 3.681,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042956.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508042959.972, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042959.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508042961.061, "dur": 151.264,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508042961.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508043112.485, "dur": 7.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043112.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508043119.749, "dur": 4.767,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043119.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508043124.645, "dur": 1.503,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043124.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508043126.245, "dur": 11.871,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043126.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508043138.212, "dur": 6.721,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043138.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508043145.061, "dur": 1.759,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043145.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508043146.917, "dur": 8.608,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043146.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508043155.653, "dur": 10.656,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043155.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508043166.437, "dur": 6.176,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043166.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043172.740, "dur": 22.465,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043172.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508043195.333, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043195.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043198.629, "dur": 26.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043198.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508043225.669, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043225.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508043227.300, "dur": 22.433,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043227.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508043249.828, "dur": 6.817,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043249.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508043256.740, "dur": 1.249,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043256.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043258.116, "dur": 23.297,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043258.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508043281.509, "dur": 3.136,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043281.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043284.741, "dur": 29.791,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043284.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508043314.629, "dur": 1.600,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043314.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508043316.325, "dur": 31.808,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043316.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508043351.045, "dur": 1.152,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043351.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043357.092, "dur": 50.464,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043357.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043407.685, "dur": 48.672,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043407.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508043456.453, "dur": 2.048,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043456.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508043458.661, "dur": 20.736,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043458.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508043479.557, "dur": 6.624,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043479.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508043486.277, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043486.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043487.653, "dur": 13.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043487.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508043501.029, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043501.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043504.037, "dur": 17.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043504.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508043521.445, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043521.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508043522.788, "dur": 9.153,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043522.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508043532.069, "dur": 2.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043532.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508043535.205, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043535.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508043536.389, "dur": 15.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043536.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508043551.653, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043551.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043552.805, "dur": 13.760,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043552.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508043566.693, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043566.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043569.061, "dur": 3.648,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043569.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508043572.837, "dur": 0.928,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043572.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508043573.925, "dur": 168.160,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043573.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508043742.277, "dur": 7.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043742.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508043749.413, "dur": 4.832,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043749.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508043754.341, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043754.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508043755.973, "dur": 12.320,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043755.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508043768.389, "dur": 6.816,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043768.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508043775.333, "dur": 1.728,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043775.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508043777.189, "dur": 7.744,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043777.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508043785.029, "dur": 9.824,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043785.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508043794.981, "dur": 6.336,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043794.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043801.445, "dur": 22.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043801.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508043824.549, "dur": 3.200,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043824.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043827.877, "dur": 27.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043827.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508043855.045, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043855.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508043856.677, "dur": 24.320,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043856.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508043881.093, "dur": 6.688,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043881.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508043887.909, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043887.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043889.253, "dur": 23.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043889.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508043912.389, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043912.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043915.589, "dur": 29.696,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043915.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508043945.413, "dur": 1.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043945.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508043947.141, "dur": 31.552,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043947.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508043981.893, "dur": 1.152,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043981.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508043987.845, "dur": 51.392,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508043987.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044039.365, "dur": 48.416,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044039.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508044087.909, "dur": 2.048,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044087.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508044090.117, "dur": 20.928,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044090.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508044111.141, "dur": 6.752,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044111.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044118.021, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044118.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044119.365, "dur": 13.408,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044119.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508044132.869, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044132.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044135.877, "dur": 16.832,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044135.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508044152.837, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044152.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508044154.213, "dur": 9.344,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044154.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508044163.685, "dur": 2.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044163.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508044166.757, "dur": 1.056,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044166.757,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508044167.941, "dur": 14.816,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044167.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508044182.853, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044182.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044184.517, "dur": 13.888,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044184.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508044198.533, "dur": 2.240,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044198.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044200.901, "dur": 3.648,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044200.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508044204.709, "dur": 0.928,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044204.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508044205.765, "dur": 154.656,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044205.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508044360.549, "dur": 7.008,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044360.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508044367.685, "dur": 5.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044367.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508044373.029, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044373.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508044374.661, "dur": 11.680,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044374.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508044386.469, "dur": 6.976,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044386.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508044393.541, "dur": 1.728,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044393.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508044395.365, "dur": 7.520,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044395.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508044403.013, "dur": 9.440,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044403.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508044412.581, "dur": 6.080,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044412.581,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044418.789, "dur": 22.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044418.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508044441.765, "dur": 3.200,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044441.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044445.061, "dur": 27.264,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044445.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508044472.421, "dur": 1.504,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044472.421,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508044474.085, "dur": 23.552,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044474.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508044497.765, "dur": 7.040,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044497.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044504.933, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044504.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508044506.277, "dur": 19.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044506.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508044525.477, "dur": 13.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044525.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044539.173, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044539.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044540.517, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044540.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508044541.893, "dur": 26.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044541.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044568.101, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044568.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044569.445, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044569.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508044570.821, "dur": 1.024,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044570.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508044572.005, "dur": 38.080,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044572.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508044610.213, "dur": 12.096,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044610.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044622.181, "dur": 1.248,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044622.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044623.557, "dur": 10.656,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044623.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508044634.341, "dur": 2.880,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044634.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044637.349, "dur": 11.936,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044637.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508044649.381, "dur": 1.184,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044649.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508044650.661, "dur": 7.520,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044650.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044658.277, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044658.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508044659.685, "dur": 4.544,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044659.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508044664.357, "dur": 1.216,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044664.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044665.733, "dur": 10.144,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044665.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508044676.037, "dur": 2.816,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044676.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044679.013, "dur": 10.752,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044679.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508044689.925, "dur": 1.152,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044689.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290508044691.205, "dur": 2.560,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044691.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290508044693.893, "dur": 8.096,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044693.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044702.117, "dur": 22.816,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044702.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508044725.061, "dur": 3.104,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044725.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044728.293, "dur": 29.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044728.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508044757.957, "dur": 1.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044757.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508044759.653, "dur": 31.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044759.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508044793.861, "dur": 1.024,
    "args": {
      "External id": 3618,      "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044793.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044799.749, "dur": 56.320,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044799.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044856.197, "dur": 48.736,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044856.197,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508044905.061, "dur": 2.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044905.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290508044907.301, "dur": 34.048,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044907.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044941.477, "dur": 17.312,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044941.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508044958.917, "dur": 9.824,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044958.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508044968.869, "dur": 16.096,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044968.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508044985.093, "dur": 5.760,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044985.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508044990.981, "dur": 159.168,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508044990.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508045150.309, "dur": 13.632,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045150.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508045164.037, "dur": 7.360,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045164.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508045171.493, "dur": 6.080,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045171.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508045177.733, "dur": 4.864,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045177.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508045182.725, "dur": 5.664,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045182.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045188.517, "dur": 27.904,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045188.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290508045216.549, "dur": 20.608,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045216.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508045237.285, "dur": 1.280,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045237.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045238.693, "dur": 13.888,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045238.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508045252.709, "dur": 2.848,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045252.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508045255.685, "dur": 1.888,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045255.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508045257.701, "dur": 3.328,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045257.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508045261.157, "dur": 1.088,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045261.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508045262.341, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045262.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045263.973, "dur": 13.728,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045263.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508045277.829, "dur": 2.112,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045277.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508045280.101, "dur": 0.960,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045280.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508045281.189, "dur": 4.896,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045281.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508045286.213, "dur": 1.568,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045286.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508045287.877, "dur": 7.456,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045287.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508045295.429, "dur": 1.760,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045295.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508045297.285, "dur": 22.208,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045297.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508045319.621, "dur": 3.072,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045319.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508045322.821, "dur": 1.536,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045322.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290508045324.453, "dur": 24.896,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045324.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508045349.477, "dur": 1.440,
    "args": {
      "External id": 3618,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506140,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 506140, "pid": 0, "tid": 7, "ts": 1290508045349.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507993655.765, "dur": 806.120,
    "args": {
      "External id": 3618,      "cbid": 311, "correlation": 506140
    }
  },
  {
    "ph": "s", "id": 506140, "pid": 14565, "tid": 14640, "ts": 1290507993655.765,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290507997106.427, "dur": 1.035,
    "args": {
      "External id": 4386,      "cbid": 317, "correlation": 506915
    }
  },
  {
    "ph": "f", "id": 506915, "pid": 14565, "tid": 14640, "ts": 1290507997106.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290508045353.445, "dur": 8.544,
    "args": {
      "External id": 4386,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506917,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506917, "pid": 0, "tid": 7, "ts": 1290508045353.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290508045362.309, "dur": 15.744,
    "args": {
      "External id": 4386,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506917,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506917, "pid": 0, "tid": 7, "ts": 1290508045362.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290507997108.078, "dur": 12.150,
    "args": {
      "External id": 4386,      "cbid": 311, "correlation": 506917
    }
  },
  {
    "ph": "s", "id": 506917, "pid": 14565, "tid": 14640, "ts": 1290507997108.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290508045379.333, "dur": 22.208,
    "args": {
      "External id": 4388,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 506928,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 506928, "pid": 0, "tid": 7, "ts": 1290508045379.333,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290507997164.404, "dur": 12.084,
    "args": {
      "External id": 4388,      "cbid": 211, "correlation": 506928
    }
  },
  {
    "ph": "s", "id": 506928, "pid": 14565, "tid": 14640, "ts": 1290507997164.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290508051838.694, "dur": 2.016,
    "args": {
      "External id": 4396,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507013,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507013, "pid": 0, "tid": 7, "ts": 1290508051838.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290508000629.911, "dur": 19.021,
    "args": {
      "External id": 4396,      "cbid": 211, "correlation": 507013
    }
  },
  {
    "ph": "s", "id": 507013, "pid": 14565, "tid": 14640, "ts": 1290508000629.911,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290508000657.944, "dur": 0.721,
    "args": {
      "External id": 4394,      "cbid": 317, "correlation": 507021
    }
  },
  {
    "ph": "f", "id": 507021, "pid": 14565, "tid": 14640, "ts": 1290508000657.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290508051844.966, "dur": 12.256,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051844.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290508051857.318, "dur": 18.144,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051857.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290508051875.558, "dur": 6.592,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051875.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508051882.246, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051882.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508051883.622, "dur": 23.328,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051883.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508051907.078, "dur": 3.200,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051907.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508051910.406, "dur": 30.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051910.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508051940.774, "dur": 1.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051940.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508051942.534, "dur": 32.736,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051942.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508051978.310, "dur": 1.024,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051978.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508051984.262, "dur": 56.544,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508051984.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052040.934, "dur": 48.672,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052040.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508052089.702, "dur": 2.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052089.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290508052091.910, "dur": 29.120,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052091.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508052121.158, "dur": 6.432,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052121.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508052127.718, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052127.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052129.094, "dur": 13.440,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052129.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508052142.663, "dur": 2.879,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052142.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052145.639, "dur": 16.256,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052145.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508052162.055, "dur": 1.215,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052162.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508052163.398, "dur": 9.728,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052163.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508052173.286, "dur": 3.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052173.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508052176.455, "dur": 1.119,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052176.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508052177.671, "dur": 14.943,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052177.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508052192.742, "dur": 1.025,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052192.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052193.894, "dur": 13.824,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052193.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508052207.847, "dur": 2.239,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052207.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052210.214, "dur": 3.680,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052210.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508052214.054, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052214.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508052215.142, "dur": 155.360,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052215.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508052370.631, "dur": 7.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052370.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508052377.926, "dur": 4.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052377.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508052382.982, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052382.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508052384.614, "dur": 12.193,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052384.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508052396.934, "dur": 6.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052396.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508052404.039, "dur": 1.663,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052404.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508052405.830, "dur": 7.520,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052405.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508052413.478, "dur": 8.864,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052413.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508052422.438, "dur": 5.888,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052422.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052428.454, "dur": 22.464,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052428.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508052451.047, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052451.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052454.278, "dur": 26.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052454.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508052481.383, "dur": 1.503,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052481.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508052483.014, "dur": 20.512,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052483.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508052503.655, "dur": 7.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052503.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508052510.790, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052510.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052512.166, "dur": 23.457,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052512.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508052535.750, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052535.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052538.983, "dur": 29.759,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052538.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508052568.839, "dur": 1.599,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052568.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508052570.566, "dur": 32.384,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052570.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508052604.998, "dur": 1.121,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.12845673505798394
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052604.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052610.951, "dur": 49.759,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052610.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052660.839, "dur": 48.319,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052660.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508052709.287, "dur": 2.111,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052709.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508052711.527, "dur": 21.888,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052711.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508052733.542, "dur": 6.625,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052733.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508052740.263, "dur": 1.247,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052740.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052741.639, "dur": 13.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052741.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508052755.047, "dur": 2.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052755.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052758.023, "dur": 16.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052758.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508052775.078, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052775.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508052776.455, "dur": 9.664,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052776.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508052786.246, "dur": 2.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052786.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508052789.351, "dur": 1.087,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052789.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508052790.567, "dur": 14.975,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052790.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508052805.671, "dur": 1.023,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052805.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052806.791, "dur": 13.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052806.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508052820.518, "dur": 2.272,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052820.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508052822.887, "dur": 3.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052822.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508052826.695, "dur": 0.927,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052826.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508052827.751, "dur": 158.528,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052827.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508052986.439, "dur": 7.328,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052986.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508052993.895, "dur": 4.864,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052993.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508052998.854, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508052998.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508053000.519, "dur": 12.031,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053000.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508053012.679, "dur": 6.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053012.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508053019.782, "dur": 1.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053019.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508053021.607, "dur": 7.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053021.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508053029.383, "dur": 10.047,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053029.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508053039.559, "dur": 6.143,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053039.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053045.831, "dur": 22.719,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053045.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508053068.679, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053068.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053071.911, "dur": 27.200,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053071.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508053099.206, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053099.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508053100.871, "dur": 25.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053100.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508053126.023, "dur": 6.688,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053126.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508053132.838, "dur": 1.249,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053132.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053134.214, "dur": 22.913,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053134.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508053157.255, "dur": 3.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053157.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053160.519, "dur": 30.144,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053160.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508053190.790, "dur": 1.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053190.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508053192.519, "dur": 31.328,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053192.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508053225.478, "dur": 1.057,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.13623462630085148
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053225.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053231.399, "dur": 50.464,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053231.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053281.991, "dur": 48.064,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053281.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508053330.183, "dur": 2.112,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053330.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508053332.455, "dur": 21.472,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053332.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508053354.055, "dur": 6.592,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053354.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508053360.775, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053360.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053362.118, "dur": 13.537,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053362.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508053375.783, "dur": 2.911,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053375.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053378.791, "dur": 16.479,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053378.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508053395.399, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053395.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508053396.743, "dur": 9.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053396.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508053406.471, "dur": 2.752,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053406.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508053409.383, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053409.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508053410.567, "dur": 14.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053410.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508053425.606, "dur": 1.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053425.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053427.399, "dur": 13.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053427.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508053441.159, "dur": 2.208,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053441.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053443.495, "dur": 3.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053443.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508053447.239, "dur": 0.992,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053447.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508053448.327, "dur": 161.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053448.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508053609.543, "dur": 7.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053609.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508053616.743, "dur": 4.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053616.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508053621.831, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053621.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508053623.463, "dur": 11.840,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053623.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508053635.431, "dur": 6.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053635.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508053642.503, "dur": 1.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053642.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508053644.327, "dur": 7.424,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053644.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508053651.879, "dur": 8.512,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053651.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508053660.487, "dur": 6.112,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053660.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053666.727, "dur": 23.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053666.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508053689.927, "dur": 3.200,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053689.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053693.223, "dur": 27.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053693.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508053720.391, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053720.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508053722.055, "dur": 24.864,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053722.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508053747.047, "dur": 6.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053747.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508053753.991, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053753.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053755.367, "dur": 23.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053755.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508053778.631, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053778.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053781.895, "dur": 30.048,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053781.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508053812.071, "dur": 1.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053812.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508053813.831, "dur": 31.808,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053813.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508053848.135, "dur": 0.992,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053848.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053853.895, "dur": 50.592,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053853.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053904.615, "dur": 48.320,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053904.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508053953.063, "dur": 2.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053953.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508053955.271, "dur": 21.408,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053955.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508053976.807, "dur": 6.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053976.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508053983.751, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053983.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508053985.095, "dur": 13.344,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053985.095,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508053998.567, "dur": 2.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508053998.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054001.575, "dur": 15.808,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054001.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508054017.511, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054017.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508054018.919, "dur": 9.152,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054018.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508054028.199, "dur": 2.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054028.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508054031.271, "dur": 1.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054031.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508054032.423, "dur": 14.752,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054032.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508054047.271, "dur": 1.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054047.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054048.391, "dur": 13.408,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054048.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508054061.927, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054061.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054064.295, "dur": 3.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054064.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508054068.071, "dur": 0.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054068.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508054069.127, "dur": 158.176,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054069.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508054227.495, "dur": 6.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054227.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508054234.599, "dur": 4.800,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054234.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508054239.527, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054239.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508054241.159, "dur": 11.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054241.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508054252.903, "dur": 7.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054252.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508054260.039, "dur": 1.664,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054260.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508054261.831, "dur": 7.872,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054261.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508054269.831, "dur": 9.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054269.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508054279.271, "dur": 5.856,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054279.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054285.223, "dur": 21.920,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054285.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508054307.303, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054307.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054310.567, "dur": 26.944,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054310.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508054337.639, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054337.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508054339.271, "dur": 23.296,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054339.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508054362.695, "dur": 6.784,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054362.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508054369.575, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054369.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054370.951, "dur": 22.944,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054370.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508054394.023, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054394.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054397.255, "dur": 29.888,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054397.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508054427.271, "dur": 1.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054427.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508054428.999, "dur": 31.744,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054428.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508054462.407, "dur": 1.024,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054462.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054468.327, "dur": 52.192,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054468.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054520.615, "dur": 48.576,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054520.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508054569.319, "dur": 2.048,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054569.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508054571.495, "dur": 20.576,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054571.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508054592.199, "dur": 6.624,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054592.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508054598.919, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054598.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054600.295, "dur": 13.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054600.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508054613.479, "dur": 2.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054613.479,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054616.455, "dur": 17.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054616.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508054633.671, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054633.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508054635.047, "dur": 10.016,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054635.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508054645.191, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054645.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508054648.423, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054648.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508054649.607, "dur": 15.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054649.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508054664.903, "dur": 1.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054664.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054666.055, "dur": 14.144,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054666.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508054680.295, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054680.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054682.663, "dur": 3.680,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054682.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508054686.439, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054686.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508054687.527, "dur": 164.544,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054687.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508054852.231, "dur": 7.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054852.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508054859.367, "dur": 4.896,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054859.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508054864.391, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054864.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508054866.023, "dur": 11.584,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054866.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508054877.703, "dur": 7.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054877.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508054884.839, "dur": 1.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054884.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508054886.663, "dur": 7.488,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054886.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508054894.311, "dur": 9.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054894.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508054903.495, "dur": 6.144,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054903.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054909.735, "dur": 22.944,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054909.735,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508054932.807, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054932.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508054936.071, "dur": 26.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054936.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508054963.015, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054963.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508054964.647, "dur": 28.320,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054964.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508054993.063, "dur": 6.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508054993.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508055000.007, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055000.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055001.351, "dur": 23.520,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055001.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508055025.031, "dur": 3.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055025.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055028.199, "dur": 29.984,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055028.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508055058.311, "dur": 1.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055058.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508055060.039, "dur": 31.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055060.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508055093.159, "dur": 1.024,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055093.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055099.111, "dur": 51.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055099.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055150.279, "dur": 47.776,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055150.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508055198.183, "dur": 2.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055198.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508055200.391, "dur": 21.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055200.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508055221.543, "dur": 6.624,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055221.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508055228.295, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055228.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055229.639, "dur": 13.152,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055229.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508055242.887, "dur": 2.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055242.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055245.831, "dur": 16.640,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055245.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508055262.599, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055262.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508055263.975, "dur": 8.896,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055263.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508055272.999, "dur": 3.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055272.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508055276.167, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055276.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508055277.319, "dur": 15.360,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055277.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508055292.775, "dur": 1.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055292.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055293.927, "dur": 13.120,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055293.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508055307.143, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055307.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055309.511, "dur": 3.744,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055309.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508055313.383, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055313.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508055314.471, "dur": 153.184,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055314.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508055467.783, "dur": 7.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055467.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508055475.079, "dur": 4.800,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055475.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508055480.039, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055480.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508055481.671, "dur": 12.096,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055481.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508055493.895, "dur": 6.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055493.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508055500.999, "dur": 1.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055500.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508055502.823, "dur": 7.520,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055502.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508055510.471, "dur": 8.704,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055510.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508055519.303, "dur": 5.824,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055519.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055525.223, "dur": 22.944,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055525.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508055548.263, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055548.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055551.495, "dur": 27.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055551.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508055578.759, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055578.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508055580.391, "dur": 25.760,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055580.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508055606.279, "dur": 6.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055606.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508055613.223, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055613.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055614.599, "dur": 23.232,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055614.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508055637.959, "dur": 3.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055637.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055641.159, "dur": 29.792,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055641.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508055671.111, "dur": 1.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055671.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508055672.871, "dur": 31.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055672.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508055706.663, "dur": 1.120,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055706.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055712.647, "dur": 50.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055712.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055762.983, "dur": 48.608,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055762.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508055811.687, "dur": 2.144,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055811.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508055813.927, "dur": 20.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055813.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508055834.983, "dur": 6.560,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055834.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508055841.671, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055841.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055843.015, "dur": 13.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055843.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508055856.391, "dur": 2.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055856.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055859.367, "dur": 16.448,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055859.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508055875.943, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055875.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508055877.319, "dur": 9.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055877.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508055886.503, "dur": 3.296,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055886.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508055889.927, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055889.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508055891.111, "dur": 15.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055891.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508055906.343, "dur": 0.992,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055906.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055907.463, "dur": 13.760,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055907.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508055921.319, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055921.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508055923.687, "dur": 3.680,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055923.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508055927.463, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055927.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508055928.551, "dur": 156.672,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508055928.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508056085.383, "dur": 7.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056085.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508056092.583, "dur": 4.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056092.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508056097.639, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056097.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508056099.303, "dur": 11.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056099.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508056111.047, "dur": 7.328,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056111.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508056118.471, "dur": 1.728,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056118.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508056120.327, "dur": 7.552,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056120.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508056128.007, "dur": 10.016,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056128.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508056138.119, "dur": 6.112,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056138.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056144.327, "dur": 22.528,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056144.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508056166.983, "dur": 3.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056166.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056170.183, "dur": 27.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056170.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508056197.447, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056197.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508056199.079, "dur": 24.608,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056199.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508056223.847, "dur": 6.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056223.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056230.855, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056230.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056232.199, "dur": 23.456,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056232.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508056255.783, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056255.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056258.983, "dur": 30.208,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056258.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508056289.319, "dur": 1.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056289.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508056291.047, "dur": 32.544,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056291.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508056325.159, "dur": 1.024,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056325.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056331.111, "dur": 52.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056331.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056384.199, "dur": 48.096,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056384.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508056432.423, "dur": 2.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056432.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508056434.631, "dur": 21.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056434.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508056456.071, "dur": 6.592,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056456.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056462.791, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056462.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056464.135, "dur": 13.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056464.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508056477.511, "dur": 2.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056477.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056480.519, "dur": 16.768,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056480.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508056497.447, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056497.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508056498.791, "dur": 9.152,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056498.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508056508.071, "dur": 2.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056508.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508056511.111, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056511.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508056512.295, "dur": 15.904,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056512.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508056528.327, "dur": 1.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056528.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056529.543, "dur": 13.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056529.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508056543.271, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056543.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056545.639, "dur": 3.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056545.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508056549.415, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056549.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508056550.503, "dur": 163.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056550.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508056713.799, "dur": 7.328,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056713.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508056721.223, "dur": 4.896,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056721.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508056726.215, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056726.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508056727.879, "dur": 11.552,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056727.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508056739.559, "dur": 6.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056739.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508056746.631, "dur": 1.664,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056746.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508056748.455, "dur": 7.488,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056748.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508056756.071, "dur": 9.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056756.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508056765.319, "dur": 6.016,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056765.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056771.463, "dur": 22.784,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056771.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508056794.343, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056794.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056797.607, "dur": 27.232,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056797.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508056824.967, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056824.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508056826.599, "dur": 20.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056826.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508056847.687, "dur": 7.232,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056847.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056855.047, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056855.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508056856.391, "dur": 19.264,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056856.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508056875.751, "dur": 13.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056875.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056889.447, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056889.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056890.855, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056890.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508056892.199, "dur": 25.664,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056892.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056917.991, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056917.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056919.335, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056919.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508056920.679, "dur": 1.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056920.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508056921.863, "dur": 37.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056921.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508056959.207, "dur": 10.209,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056959.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508056969.543, "dur": 1.824,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056969.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056971.495, "dur": 10.528,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056971.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508056982.151, "dur": 2.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056982.151,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508056985.159, "dur": 11.777,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056985.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508056997.031, "dur": 1.152,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056997.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508056998.343, "dur": 7.552,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508056998.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508057005.991, "dur": 1.249,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057005.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508057007.367, "dur": 4.480,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057007.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508057011.975, "dur": 1.217,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057011.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057013.319, "dur": 10.272,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057013.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057023.687, "dur": 2.945,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057023.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057026.727, "dur": 10.624,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057026.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508057037.479, "dur": 1.152,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057037.479,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290508057038.759, "dur": 2.496,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057038.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290508057041.383, "dur": 7.424,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057041.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057048.903, "dur": 22.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057048.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057071.911, "dur": 3.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057071.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057075.111, "dur": 30.336,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057075.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508057105.576, "dur": 1.599,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057105.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508057107.303, "dur": 31.744,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057107.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508057141.287, "dur": 1.088,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057141.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057147.207, "dur": 53.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057147.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057200.936, "dur": 48.575,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057200.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508057249.671, "dur": 2.049,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057249.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508057251.847, "dur": 29.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057251.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508057281.671, "dur": 6.817,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057281.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508057288.584, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057288.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057289.959, "dur": 13.505,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057289.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057303.591, "dur": 2.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057303.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057306.535, "dur": 16.192,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057306.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508057322.856, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057322.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508057324.263, "dur": 9.120,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057324.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508057333.511, "dur": 3.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057333.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508057336.679, "dur": 1.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057336.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508057337.896, "dur": 15.167,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057337.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508057353.192, "dur": 1.471,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057353.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057354.791, "dur": 13.824,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057354.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057368.743, "dur": 2.272,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057368.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057371.144, "dur": 3.647,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057371.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508057374.920, "dur": 0.991,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057374.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508057376.040, "dur": 154.048,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057376.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508057530.215, "dur": 6.945,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057530.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508057537.287, "dur": 4.896,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057537.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508057542.312, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057542.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508057543.943, "dur": 12.353,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057543.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508057556.392, "dur": 6.815,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057556.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508057563.304, "dur": 1.759,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057563.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508057565.191, "dur": 7.424,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057565.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508057572.743, "dur": 9.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057572.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508057581.896, "dur": 6.112,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057581.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057588.135, "dur": 22.624,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057588.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057610.887, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057610.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057614.151, "dur": 27.264,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057614.151,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508057641.575, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057641.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508057643.239, "dur": 23.553,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057643.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508057666.919, "dur": 6.785,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057666.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508057673.800, "dur": 1.247,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057673.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057675.176, "dur": 23.359,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057675.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057698.664, "dur": 3.167,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057698.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057701.927, "dur": 29.952,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057701.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508057732.007, "dur": 1.665,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057732.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508057733.799, "dur": 31.712,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057733.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508057767.911, "dur": 1.025,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.1404878048780488
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057767.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057773.736, "dur": 50.560,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057773.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057824.423, "dur": 48.385,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057824.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508057872.968, "dur": 2.047,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057872.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508057875.144, "dur": 20.608,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057875.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508057895.848, "dur": 6.656,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057895.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508057902.631, "dur": 1.249,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057902.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057903.976, "dur": 13.151,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057903.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057917.256, "dur": 2.847,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057917.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057920.232, "dur": 16.831,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057920.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508057937.192, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057937.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508057938.568, "dur": 9.184,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057938.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508057947.880, "dur": 3.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057947.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508057951.016, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057951.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508057952.199, "dur": 14.849,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057952.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508057967.175, "dur": 1.729,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057967.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057969.031, "dur": 13.824,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057969.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508057982.984, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057982.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508057985.351, "dur": 3.649,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057985.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508057989.127, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057989.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508057990.216, "dur": 158.655,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508057990.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508058149.031, "dur": 6.945,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058149.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508058156.104, "dur": 4.864,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058156.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508058161.064, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058161.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508058162.696, "dur": 11.712,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058162.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508058174.504, "dur": 6.879,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058174.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508058181.512, "dur": 1.664,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058181.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508058183.304, "dur": 7.936,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058183.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508058191.400, "dur": 9.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058191.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508058200.584, "dur": 6.176,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058200.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058206.920, "dur": 23.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058206.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508058230.696, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058230.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058233.960, "dur": 27.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058233.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508058261.128, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058261.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508058262.760, "dur": 25.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058262.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508058288.200, "dur": 6.975,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058288.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508058295.272, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058295.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058296.648, "dur": 23.808,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058296.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508058320.552, "dur": 3.232,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058320.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058323.880, "dur": 29.824,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058323.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508058353.832, "dur": 1.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058353.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508058355.560, "dur": 31.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058355.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508058388.488, "dur": 1.024,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058388.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058394.280, "dur": 52.096,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058394.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058446.504, "dur": 48.416,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058446.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508058495.080, "dur": 2.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058495.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508058497.288, "dur": 21.856,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058497.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508058519.272, "dur": 6.752,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058519.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508058526.120, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058526.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058527.496, "dur": 13.184,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058527.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508058540.808, "dur": 2.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058540.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058543.784, "dur": 15.744,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058543.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508058559.656, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058559.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508058561.032, "dur": 9.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058561.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508058570.184, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058570.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508058573.448, "dur": 1.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058573.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508058574.632, "dur": 15.424,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058574.632,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508058590.184, "dur": 0.992,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058590.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058591.304, "dur": 13.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058591.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508058605.032, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058605.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058607.400, "dur": 3.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058607.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508058611.176, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058611.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508058612.232, "dur": 157.760,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058612.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508058770.088, "dur": 7.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058770.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508058777.288, "dur": 4.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058777.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508058782.376, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058782.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508058784.008, "dur": 11.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058784.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508058795.816, "dur": 6.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058795.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508058802.856, "dur": 1.792,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058802.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508058804.776, "dur": 8.352,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058804.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508058813.224, "dur": 10.784,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058813.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508058824.168, "dur": 6.208,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058824.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058830.504, "dur": 23.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058830.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508058853.800, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058853.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058857.032, "dur": 27.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058857.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508058884.200, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058884.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508058885.864, "dur": 22.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058885.864,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508058908.936, "dur": 6.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058908.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508058915.944, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058915.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058917.288, "dur": 23.200,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058917.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508058940.616, "dur": 3.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058940.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508058943.816, "dur": 30.176,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058943.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508058974.120, "dur": 1.632,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058974.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508058975.880, "dur": 31.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508058975.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508059009.096, "dur": 1.088,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059009.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059015.016, "dur": 49.376,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059015.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059064.520, "dur": 48.384,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059064.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508059113.032, "dur": 2.048,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059113.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508059115.208, "dur": 21.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059115.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508059136.616, "dur": 6.624,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059136.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508059143.400, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059143.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059144.744, "dur": 13.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059144.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508059158.088, "dur": 2.944,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059158.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059161.128, "dur": 16.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059161.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508059178.184, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059178.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508059179.592, "dur": 9.408,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059179.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508059189.128, "dur": 3.264,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059189.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508059192.520, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059192.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508059193.704, "dur": 14.368,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059193.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508059208.200, "dur": 0.992,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059208.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059209.320, "dur": 13.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059209.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508059223.048, "dur": 2.208,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059223.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059225.384, "dur": 3.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059225.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508059229.128, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059229.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508059230.184, "dur": 167.840,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059230.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508059398.184, "dur": 7.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059398.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508059405.448, "dur": 4.800,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059405.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508059410.376, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059410.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508059412.008, "dur": 12.992,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059412.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508059425.096, "dur": 6.784,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059425.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508059431.976, "dur": 1.792,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059431.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508059433.864, "dur": 7.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059433.864,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508059441.576, "dur": 8.608,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059441.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508059450.280, "dur": 5.888,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059450.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059456.328, "dur": 22.528,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059456.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508059478.984, "dur": 3.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059478.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059482.152, "dur": 27.232,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059482.152,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508059509.512, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059509.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508059511.144, "dur": 23.552,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059511.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508059534.824, "dur": 6.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059534.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508059541.768, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059541.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059543.112, "dur": 23.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059543.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508059566.312, "dur": 3.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059566.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059569.512, "dur": 29.856,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059569.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508059599.464, "dur": 1.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059599.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508059601.256, "dur": 31.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059601.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508059634.728, "dur": 1.056,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059634.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059640.616, "dur": 51.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059640.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059691.912, "dur": 48.576,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059691.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508059740.616, "dur": 2.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059740.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508059742.824, "dur": 21.344,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059742.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508059764.296, "dur": 6.624,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059764.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508059771.048, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059771.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059772.392, "dur": 13.344,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059772.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508059785.832, "dur": 2.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059785.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059788.840, "dur": 16.768,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059788.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508059805.736, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059805.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508059807.080, "dur": 9.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059807.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508059816.680, "dur": 3.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059816.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508059819.880, "dur": 1.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059819.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508059821.064, "dur": 15.648,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059821.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508059836.840, "dur": 1.376,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059836.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059838.376, "dur": 13.792,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059838.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508059852.328, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059852.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508059854.696, "dur": 3.584,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059854.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508059858.440, "dur": 0.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059858.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508059859.496, "dur": 158.304,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508059859.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508060017.928, "dur": 7.232,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060017.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508060025.288, "dur": 4.768,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060025.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508060030.152, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060030.152,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508060031.784, "dur": 11.872,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060031.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508060043.752, "dur": 6.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060043.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508060050.696, "dur": 1.728,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060050.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508060052.520, "dur": 9.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060052.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508060061.736, "dur": 10.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060061.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508060072.776, "dur": 6.048,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060072.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060078.952, "dur": 22.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060078.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508060101.928, "dur": 3.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060101.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060105.160, "dur": 27.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060105.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508060132.328, "dur": 1.568,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060132.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508060134.024, "dur": 22.496,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060134.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508060156.648, "dur": 6.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060156.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508060163.656, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060163.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060165.032, "dur": 22.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060165.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508060188.008, "dur": 3.040,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060188.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060191.176, "dur": 29.984,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060191.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508060221.288, "dur": 1.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060221.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508060222.984, "dur": 31.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060222.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508060257.320, "dur": 1.024,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060257.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060263.208, "dur": 50.016,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060263.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060313.352, "dur": 48.544,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060313.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508060362.056, "dur": 2.016,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060362.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508060364.232, "dur": 21.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060364.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508060385.672, "dur": 6.656,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060385.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508060392.456, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060392.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060393.800, "dur": 13.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060393.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508060407.208, "dur": 2.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060407.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060410.216, "dur": 16.704,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060410.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508060427.048, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060427.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508060428.392, "dur": 9.152,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060428.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508060437.672, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060437.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508060440.904, "dur": 1.088,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060440.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508060442.088, "dur": 15.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060442.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508060457.800, "dur": 1.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060457.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060458.952, "dur": 13.568,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060458.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508060472.648, "dur": 2.240,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060472.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060474.984, "dur": 3.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060474.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508060478.728, "dur": 0.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060478.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508060479.784, "dur": 159.488,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060479.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508060639.464, "dur": 7.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060639.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508060646.600, "dur": 4.800,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060646.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508060651.528, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060651.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508060653.160, "dur": 12.704,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060653.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508060665.992, "dur": 6.752,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060665.992,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508060672.872, "dur": 1.728,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060672.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508060674.728, "dur": 7.424,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060674.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508060682.280, "dur": 9.408,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060682.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508060691.848, "dur": 6.048,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060691.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060698.024, "dur": 22.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060698.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508060720.232, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060720.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060723.496, "dur": 26.784,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060723.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508060750.408, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060750.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508060752.040, "dur": 24.064,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060752.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508060776.232, "dur": 6.656,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060776.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508060782.984, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060782.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060784.360, "dur": 23.168,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060784.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508060807.624, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060807.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060810.824, "dur": 30.144,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060810.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508060841.096, "dur": 1.600,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060841.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508060842.824, "dur": 31.968,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060842.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508060877.928, "dur": 1.152,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060877.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060883.880, "dur": 52.736,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060883.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508060936.744, "dur": 47.904,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060936.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508060984.744, "dur": 2.112,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060984.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508060986.984, "dur": 20.832,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508060986.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508061007.944, "dur": 6.656,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061007.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061014.696, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061014.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061016.072, "dur": 13.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061016.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508061029.512, "dur": 2.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061029.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061032.488, "dur": 16.736,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061032.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508061049.320, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061049.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508061050.696, "dur": 9.312,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061050.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508061060.136, "dur": 2.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061060.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "overhead", "name": "Activity Buffer Request", "pid": -1, "tid": 0,
    "ts": 1290508001323.837, "dur": 1684.013
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508061063.144, "dur": 1.056,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061063.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508061064.328, "dur": 15.072,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061064.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508061079.528, "dur": 1.728,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061079.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061081.384, "dur": 13.664,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061081.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508061095.208, "dur": 2.208,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061095.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061097.544, "dur": 3.744,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061097.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508061101.416, "dur": 0.928,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061101.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508061102.472, "dur": 152.768,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061102.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508061255.368, "dur": 6.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061255.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508061262.472, "dur": 4.992,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061262.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508061267.624, "dur": 1.504,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061267.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508061269.224, "dur": 11.712,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061269.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508061281.032, "dur": 7.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061281.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508061288.168, "dur": 1.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061288.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508061289.992, "dur": 7.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061289.992,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508061297.704, "dur": 9.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061297.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508061307.048, "dur": 5.952,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061307.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061313.160, "dur": 23.296,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061313.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508061336.584, "dur": 3.104,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061336.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061339.784, "dur": 27.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061339.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508061367.048, "dur": 1.472,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061367.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508061368.680, "dur": 22.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061368.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508061391.688, "dur": 6.944,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061391.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061398.760, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061398.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508061400.136, "dur": 19.296,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061400.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508061419.528, "dur": 13.696,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061419.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061433.352, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061433.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061434.696, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061434.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508061436.072, "dur": 25.728,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061436.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061461.896, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061461.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061463.240, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061463.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508061464.616, "dur": 1.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061464.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508061465.768, "dur": 37.440,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061465.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508061503.336, "dur": 14.688,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061503.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061517.928, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061517.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061519.304, "dur": 10.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061519.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508061530.248, "dur": 2.912,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061530.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061533.256, "dur": 11.872,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061533.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508061545.256, "dur": 1.152,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061545.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508061546.536, "dur": 7.488,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061546.536,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061554.120, "dur": 1.248,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061554.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508061555.496, "dur": 4.480,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061555.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508061560.104, "dur": 1.216,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061560.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061561.448, "dur": 10.176,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061561.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508061571.752, "dur": 2.816,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061571.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061574.696, "dur": 10.560,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061574.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508061585.384, "dur": 1.184,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061585.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290508061586.696, "dur": 2.528,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061586.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290508061589.352, "dur": 8.224,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061589.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061597.704, "dur": 22.976,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061597.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508061620.808, "dur": 3.008,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061620.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061623.944, "dur": 29.856,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061623.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508061653.928, "dur": 1.568,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061653.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508061655.624, "dur": 30.880,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061655.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508061687.816, "dur": 1.088,
    "args": {
      "External id": 4394,      "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061687.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061693.736, "dur": 54.176,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061693.736,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061748.040, "dur": 48.576,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061748.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508061796.744, "dur": 2.176,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061796.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290508061799.048, "dur": 32.864,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061799.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061832.040, "dur": 16.992,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061832.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508061849.160, "dur": 9.408,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061849.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508061858.696, "dur": 15.840,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061858.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508061874.664, "dur": 5.024,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061874.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508061879.816, "dur": 160.193,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508061879.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508062040.200, "dur": 13.760,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062040.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508062054.056, "dur": 7.392,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062054.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508062061.577, "dur": 6.079,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062061.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508062067.784, "dur": 4.897,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062067.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508062072.808, "dur": 5.825,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062072.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062078.760, "dur": 27.616,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062078.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290508062106.504, "dur": 20.384,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062106.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508062127.016, "dur": 1.280,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062127.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062128.392, "dur": 14.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062128.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508062142.600, "dur": 2.848,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062142.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508062145.576, "dur": 2.176,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062145.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508062147.880, "dur": 3.488,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062147.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508062151.496, "dur": 1.089,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062151.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508062152.712, "dur": 1.376,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062152.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062154.216, "dur": 13.505,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062154.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508062167.881, "dur": 2.047,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062167.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508062170.088, "dur": 0.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062170.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508062171.144, "dur": 4.960,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062171.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508062176.200, "dur": 1.568,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062176.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508062177.865, "dur": 7.359,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062177.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508062185.320, "dur": 1.728,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062185.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508062187.144, "dur": 22.080,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062187.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508062209.321, "dur": 3.136,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062209.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508062212.584, "dur": 1.536,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062212.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290508062214.248, "dur": 24.864,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062214.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508062239.240, "dur": 1.440,
    "args": {
      "External id": 4394,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507023,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507023, "pid": 0, "tid": 7, "ts": 1290508062239.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290508000659.336, "dur": 2490.231,
    "args": {
      "External id": 4394,      "cbid": 311, "correlation": 507023
    }
  },
  {
    "ph": "s", "id": 507023, "pid": 14565, "tid": 14640, "ts": 1290508000659.336,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290508006916.155, "dur": 1.020,
    "args": {
      "External id": 5162,      "cbid": 317, "correlation": 507798
    }
  },
  {
    "ph": "f", "id": 507798, "pid": 14565, "tid": 14640, "ts": 1290508006916.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290508062244.329, "dur": 8.671,
    "args": {
      "External id": 5162,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507800,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507800, "pid": 0, "tid": 7, "ts": 1290508062244.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290508062253.352, "dur": 15.872,
    "args": {
      "External id": 5162,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507800,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507800, "pid": 0, "tid": 7, "ts": 1290508062253.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290508006917.910, "dur": 13.911,
    "args": {
      "External id": 5162,      "cbid": 311, "correlation": 507800
    }
  },
  {
    "ph": "s", "id": 507800, "pid": 14565, "tid": 14640, "ts": 1290508006917.910,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290508062270.280, "dur": 22.144,
    "args": {
      "External id": 5164,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507811,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507811, "pid": 0, "tid": 7, "ts": 1290508062270.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290508006971.100, "dur": 12.159,
    "args": {
      "External id": 5164,      "cbid": 211, "correlation": 507811
    }
  },
  {
    "ph": "s", "id": 507811, "pid": 14565, "tid": 14640, "ts": 1290508006971.100,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290508068732.778, "dur": 2.016,
    "args": {
      "External id": 5172,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507896,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507896, "pid": 0, "tid": 7, "ts": 1290508068732.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290508010185.892, "dur": 12.602,
    "args": {
      "External id": 5172,      "cbid": 211, "correlation": 507896
    }
  },
  {
    "ph": "s", "id": 507896, "pid": 14565, "tid": 14640, "ts": 1290508010185.892,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290508010206.329, "dur": 0.656,
    "args": {
      "External id": 5170,      "cbid": 317, "correlation": 507904
    }
  },
  {
    "ph": "f", "id": 507904, "pid": 14565, "tid": 14640, "ts": 1290508010206.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290508068738.922, "dur": 12.288,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068738.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290508068751.338, "dur": 17.792,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068751.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290508068769.226, "dur": 6.464,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068769.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508068775.786, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068775.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508068777.130, "dur": 23.264,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068777.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508068800.522, "dur": 3.232,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068800.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508068803.882, "dur": 29.888,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068803.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508068833.866, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068833.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508068835.626, "dur": 32.448,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068835.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508068870.090, "dur": 0.992,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068870.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508068875.914, "dur": 56.608,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068875.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508068932.650, "dur": 48.640,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068932.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508068981.386, "dur": 2.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068981.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290508068983.562, "dur": 29.440,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508068983.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508069013.130, "dur": 6.464,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069013.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508069019.722, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069019.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069021.098, "dur": 13.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069021.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508069034.474, "dur": 2.816,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069034.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069037.418, "dur": 16.096,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069037.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508069053.642, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069053.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508069055.018, "dur": 9.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069055.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508069064.426, "dur": 3.392,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069064.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508069067.914, "dur": 1.088,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069067.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508069069.098, "dur": 14.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069069.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508069084.074, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069084.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069085.226, "dur": 13.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069085.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508069098.954, "dur": 2.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069098.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069101.322, "dur": 3.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069101.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508069105.066, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069105.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508069106.154, "dur": 163.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069106.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508069269.930, "dur": 6.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069269.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508069277.034, "dur": 4.832,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069277.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508069281.994, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069281.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508069283.626, "dur": 11.584,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069283.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508069295.338, "dur": 7.360,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069295.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508069302.794, "dur": 1.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069302.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508069304.618, "dur": 7.552,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069304.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508069312.298, "dur": 8.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069312.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508069321.386, "dur": 5.920,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069321.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069327.434, "dur": 22.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069327.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508069350.506, "dur": 3.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069350.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069353.802, "dur": 27.200,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069353.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508069381.130, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069381.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508069382.762, "dur": 20.480,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069382.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508069403.338, "dur": 6.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069403.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508069410.314, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069410.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069411.690, "dur": 23.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069411.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508069434.858, "dur": 3.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069434.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069438.058, "dur": 29.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069438.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508069467.850, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069467.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508069469.578, "dur": 31.360,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069469.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508069502.986, "dur": 1.056,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069502.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069508.906, "dur": 52.032,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069508.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069561.066, "dur": 48.416,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069561.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508069609.610, "dur": 2.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069609.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508069611.850, "dur": 21.376,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069611.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508069633.354, "dur": 6.560,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069633.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508069640.010, "dur": 1.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069640.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069641.386, "dur": 13.312,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069641.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508069654.794, "dur": 2.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069654.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069657.834, "dur": 16.704,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069657.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508069674.698, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069674.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508069676.074, "dur": 9.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069676.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508069685.930, "dur": 2.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069685.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508069688.970, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069688.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508069690.154, "dur": 15.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069690.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508069705.418, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069705.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069706.570, "dur": 13.440,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069706.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508069720.106, "dur": 2.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069720.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069722.442, "dur": 3.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069722.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508069726.186, "dur": 0.928,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069726.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508069727.274, "dur": 155.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069727.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508069882.474, "dur": 7.360,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069882.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508069889.962, "dur": 4.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069889.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508069894.954, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069894.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508069896.586, "dur": 11.936,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069896.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508069908.650, "dur": 7.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069908.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508069915.914, "dur": 1.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069915.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508069917.706, "dur": 7.520,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069917.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508069925.354, "dur": 9.664,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069925.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508069935.146, "dur": 6.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069935.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069941.322, "dur": 22.016,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069941.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508069963.466, "dur": 3.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069963.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508069966.698, "dur": 26.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069966.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508069993.738, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069993.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508069995.370, "dur": 24.800,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508069995.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508070020.266, "dur": 6.688,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070020.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508070027.050, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070027.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070028.426, "dur": 23.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070028.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508070051.594, "dur": 3.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070051.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070054.762, "dur": 30.048,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070054.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508070084.938, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070084.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508070086.666, "dur": 31.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070086.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508070119.402, "dur": 1.024,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070119.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070125.290, "dur": 52.800,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070125.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070178.218, "dur": 48.320,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070178.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508070226.698, "dur": 2.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070226.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508070228.906, "dur": 21.632,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070228.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508070250.666, "dur": 6.560,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070250.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508070257.354, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070257.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070258.698, "dur": 13.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070258.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508070272.106, "dur": 2.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070272.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070275.146, "dur": 16.448,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070275.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508070291.722, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070291.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508070293.066, "dur": 9.088,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070293.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508070302.282, "dur": 2.720,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070302.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508070305.162, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070305.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508070306.346, "dur": 15.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070306.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508070321.482, "dur": 1.568,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070321.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070323.146, "dur": 13.472,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070323.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508070336.746, "dur": 2.208,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070336.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070339.082, "dur": 3.648,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070339.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508070342.858, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070342.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508070343.914, "dur": 165.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070343.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508070509.098, "dur": 6.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070509.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508070516.202, "dur": 4.992,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070516.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508070521.322, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070521.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508070522.954, "dur": 12.032,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070522.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508070535.114, "dur": 7.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070535.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508070542.314, "dur": 1.824,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070542.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508070544.266, "dur": 7.392,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070544.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508070551.754, "dur": 8.384,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070551.754,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508070560.266, "dur": 5.984,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070560.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070566.346, "dur": 22.560,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070566.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508070589.002, "dur": 3.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070589.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070592.266, "dur": 26.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070592.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508070619.306, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070619.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508070620.970, "dur": 25.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070620.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508070646.826, "dur": 6.880,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070646.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508070653.834, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070653.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070655.210, "dur": 23.232,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070655.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508070678.570, "dur": 3.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070678.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070681.738, "dur": 30.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070681.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508070711.978, "dur": 1.664,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070711.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508070713.738, "dur": 32.512,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070713.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508070748.202, "dur": 1.024,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070748.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070753.994, "dur": 51.904,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070753.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070806.026, "dur": 48.480,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070806.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508070854.634, "dur": 2.048,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070854.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508070856.842, "dur": 21.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070856.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508070878.506, "dur": 6.624,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070878.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508070885.226, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070885.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070886.570, "dur": 13.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070886.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508070899.978, "dur": 2.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070899.978,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070902.986, "dur": 16.128,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070902.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508070919.242, "dur": 1.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070919.242,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508070920.682, "dur": 9.184,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070920.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508070930.026, "dur": 2.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070930.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508070933.034, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070933.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508070934.218, "dur": 14.528,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070934.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508070948.842, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070948.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070949.994, "dur": 13.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070949.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508070963.850, "dur": 2.272,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070963.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508070966.250, "dur": 3.680,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070966.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508070970.058, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070970.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508070971.146, "dur": 153.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508070971.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508071124.522, "dur": 6.880,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071124.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508071131.562, "dur": 4.896,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071131.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508071136.586, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071136.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508071138.218, "dur": 11.648,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071138.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508071149.994, "dur": 7.008,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071149.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508071157.098, "dur": 1.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071157.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508071158.922, "dur": 7.776,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071158.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508071166.826, "dur": 9.952,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071166.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508071176.874, "dur": 6.208,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071176.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071183.210, "dur": 22.784,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071183.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508071206.154, "dur": 3.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071206.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071209.418, "dur": 27.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071209.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508071236.586, "dur": 1.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071236.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508071238.250, "dur": 23.584,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071238.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508071261.930, "dur": 6.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071261.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508071268.906, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071268.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071270.250, "dur": 23.360,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071270.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508071293.738, "dur": 3.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071293.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071296.938, "dur": 29.824,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071296.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508071326.890, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071326.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508071328.618, "dur": 32.192,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071328.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508071362.602, "dur": 1.024,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071362.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071368.490, "dur": 50.016,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071368.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071418.634, "dur": 48.224,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071418.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508071466.986, "dur": 2.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071466.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508071469.194, "dur": 20.992,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071469.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508071490.314, "dur": 6.592,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071490.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508071497.034, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071497.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071498.378, "dur": 13.120,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071498.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508071511.594, "dur": 2.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071511.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071514.570, "dur": 16.736,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071514.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508071531.434, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071531.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508071532.810, "dur": 9.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071532.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508071542.698, "dur": 3.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071542.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508071545.930, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071545.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508071547.114, "dur": 14.752,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071547.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508071561.994, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071561.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071563.146, "dur": 13.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071563.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508071576.778, "dur": 2.208,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071576.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071579.114, "dur": 3.584,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071579.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508071582.826, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071582.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508071583.914, "dur": 163.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071583.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508071747.210, "dur": 6.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071747.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508071754.314, "dur": 4.832,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071754.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508071759.274, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071759.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508071760.874, "dur": 11.584,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071760.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508071772.586, "dur": 6.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071772.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508071779.658, "dur": 1.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071779.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508071781.482, "dur": 7.456,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071781.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508071789.098, "dur": 9.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071789.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508071798.250, "dur": 5.824,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071798.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071804.202, "dur": 22.592,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071804.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508071826.922, "dur": 3.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071826.922,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071830.122, "dur": 26.816,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071830.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508071857.066, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071857.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508071858.698, "dur": 28.064,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071858.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508071886.890, "dur": 6.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071886.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508071893.866, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071893.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071895.210, "dur": 22.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071895.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508071918.346, "dur": 3.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071918.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071921.579, "dur": 30.175,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071921.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508071951.882, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071951.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508071953.642, "dur": 31.712,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071953.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508071987.114, "dur": 1.184,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.12162162162162163
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071987.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508071993.194, "dur": 50.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508071993.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072044.299, "dur": 48.768,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072044.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508072093.227, "dur": 2.016,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072093.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508072095.403, "dur": 20.863,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072095.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508072116.394, "dur": 6.560,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072116.394,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508072123.083, "dur": 1.247,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072123.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072124.459, "dur": 13.503,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072124.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508072138.059, "dur": 2.879,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072138.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072141.066, "dur": 16.832,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072141.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508072158.058, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072158.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508072159.402, "dur": 9.120,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072159.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508072168.618, "dur": 2.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072168.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508072171.754, "dur": 1.057,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072171.754,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508072172.938, "dur": 15.200,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072172.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508072188.234, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072188.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072189.386, "dur": 13.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072189.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508072203.210, "dur": 2.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072203.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072205.579, "dur": 3.647,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072205.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508072209.355, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072209.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508072210.442, "dur": 166.369,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072210.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508072376.938, "dur": 7.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072376.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508072384.139, "dur": 4.832,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072384.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508072389.098, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072389.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508072390.731, "dur": 12.000,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072390.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508072402.858, "dur": 6.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072402.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508072409.963, "dur": 1.727,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072409.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508072411.786, "dur": 7.808,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072411.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508072419.690, "dur": 8.736,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072419.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508072428.522, "dur": 5.921,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072428.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072434.570, "dur": 22.496,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072434.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508072457.194, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072457.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072460.458, "dur": 27.425,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072460.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508072488.010, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072488.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508072489.643, "dur": 25.407,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072489.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508072515.178, "dur": 6.784,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072515.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508072522.090, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072522.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072523.435, "dur": 23.071,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072523.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508072546.634, "dur": 3.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072546.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072549.835, "dur": 30.047,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072549.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508072580.011, "dur": 1.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072580.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508072581.707, "dur": 32.192,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072581.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508072617.002, "dur": 1.024,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072617.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072623.051, "dur": 49.951,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072623.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072673.098, "dur": 48.608,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072673.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508072721.803, "dur": 2.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072721.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508072724.043, "dur": 21.471,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072724.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508072745.643, "dur": 6.719,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072745.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508072752.459, "dur": 1.247,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072752.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072753.835, "dur": 13.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072753.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508072767.179, "dur": 2.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072767.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072770.187, "dur": 16.416,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072770.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508072786.698, "dur": 1.249,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072786.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508072788.074, "dur": 9.473,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072788.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508072797.675, "dur": 3.423,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072797.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508072801.227, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072801.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508072802.410, "dur": 14.913,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072802.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508072817.450, "dur": 1.025,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072817.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072818.570, "dur": 13.665,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072818.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508072832.363, "dur": 2.239,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072832.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508072834.731, "dur": 3.647,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072834.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508072838.507, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072838.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508072839.563, "dur": 166.496,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508072839.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508073006.219, "dur": 7.200,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073006.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508073013.578, "dur": 4.896,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073013.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508073018.603, "dur": 1.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073018.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508073020.267, "dur": 11.520,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073020.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508073031.914, "dur": 7.265,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073031.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508073039.306, "dur": 1.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073039.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508073041.131, "dur": 7.552,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073041.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508073048.810, "dur": 9.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073048.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508073058.634, "dur": 6.272,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073058.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073065.003, "dur": 21.951,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073065.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073087.051, "dur": 3.135,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073087.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073090.315, "dur": 26.975,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073090.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508073117.450, "dur": 1.505,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073117.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508073119.083, "dur": 24.927,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073119.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508073144.107, "dur": 6.880,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073144.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073151.147, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073151.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073152.491, "dur": 22.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073152.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073175.562, "dur": 3.169,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073175.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073178.858, "dur": 30.081,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073178.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508073209.034, "dur": 1.601,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073209.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508073210.795, "dur": 31.551,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073210.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508073243.659, "dur": 1.120,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073243.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073249.643, "dur": 52.512,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073249.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073302.283, "dur": 48.512,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073302.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508073350.923, "dur": 2.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073350.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508073353.162, "dur": 21.441,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073353.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508073374.730, "dur": 6.689,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073374.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073381.515, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073381.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073382.891, "dur": 13.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073382.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073396.298, "dur": 2.849,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073396.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073399.274, "dur": 16.769,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073399.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508073416.203, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073416.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508073417.546, "dur": 8.993,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073417.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508073426.699, "dur": 2.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073426.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508073429.770, "dur": 1.057,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073429.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508073430.955, "dur": 15.328,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073430.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508073446.411, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073446.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073448.139, "dur": 13.952,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073448.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073462.219, "dur": 2.239,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073462.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073464.587, "dur": 3.680,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073464.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508073468.395, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073468.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508073469.450, "dur": 153.569,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073469.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508073623.147, "dur": 7.264,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073623.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508073630.507, "dur": 4.896,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073630.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508073635.499, "dur": 1.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073635.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508073637.163, "dur": 11.552,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073637.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508073648.843, "dur": 7.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073648.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508073656.107, "dur": 1.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073656.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508073657.963, "dur": 7.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073657.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508073665.707, "dur": 9.152,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073665.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508073674.987, "dur": 6.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073674.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073681.195, "dur": 23.392,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073681.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073704.683, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073704.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073707.915, "dur": 27.296,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073707.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508073735.307, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073735.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508073736.939, "dur": 20.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073736.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508073757.995, "dur": 7.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073757.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073765.259, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073765.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508073766.603, "dur": 18.720,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073766.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508073785.451, "dur": 13.408,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073785.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073798.987, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073798.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073800.331, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073800.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508073801.739, "dur": 25.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073801.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073827.467, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073827.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073828.811, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073828.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508073830.155, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073830.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508073831.339, "dur": 37.568,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073831.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508073869.035, "dur": 9.664,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073869.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073878.827, "dur": 2.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073878.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073881.067, "dur": 10.592,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073881.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073891.787, "dur": 2.880,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073891.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073894.795, "dur": 12.640,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073894.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508073907.563, "dur": 1.152,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073907.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508073908.811, "dur": 7.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073908.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073916.523, "dur": 1.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073916.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508073917.899, "dur": 4.448,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073917.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508073922.475, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073922.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073923.819, "dur": 10.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073923.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073934.155, "dur": 2.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073934.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073937.131, "dur": 10.496,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073937.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508073947.755, "dur": 1.152,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073947.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290508073949.035, "dur": 2.496,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073949.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290508073951.627, "dur": 7.456,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073951.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073959.179, "dur": 23.296,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073959.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508073982.635, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073982.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508073985.899, "dur": 29.312,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508073985.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508074015.339, "dur": 1.568,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074015.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508074017.035, "dur": 32.800,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074017.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508074051.595, "dur": 1.024,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074051.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074057.419, "dur": 51.968,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074057.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074109.547, "dur": 47.840,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074109.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508074157.547, "dur": 2.048,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074157.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508074159.723, "dur": 29.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074159.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508074189.611, "dur": 6.816,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074189.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508074196.523, "dur": 1.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074196.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074197.899, "dur": 13.408,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074197.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508074211.403, "dur": 2.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074211.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074214.347, "dur": 15.840,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074214.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508074230.347, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074230.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508074231.723, "dur": 9.184,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074231.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508074241.035, "dur": 3.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074241.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508074244.203, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074244.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508074245.387, "dur": 15.264,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074245.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508074260.779, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074260.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074261.931, "dur": 13.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074261.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508074275.691, "dur": 2.208,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074275.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074278.027, "dur": 3.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074278.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508074281.771, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074281.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508074282.891, "dur": 163.520,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074282.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508074446.571, "dur": 6.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074446.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508074453.611, "dur": 4.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074453.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508074458.571, "dur": 1.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074458.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508074460.203, "dur": 12.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074460.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508074473.195, "dur": 6.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074473.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508074480.171, "dur": 1.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074480.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508074482.027, "dur": 7.648,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074482.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508074489.803, "dur": 8.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074489.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508074498.795, "dur": 5.952,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074498.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074504.875, "dur": 22.656,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074504.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508074527.659, "dur": 3.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074527.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074530.859, "dur": 27.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074530.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508074558.155, "dur": 1.472,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074558.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508074559.755, "dur": 23.776,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074559.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508074583.659, "dur": 6.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074583.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508074590.667, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074590.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074592.043, "dur": 23.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074592.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508074615.211, "dur": 3.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074615.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074618.379, "dur": 29.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074618.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508074648.235, "dur": 1.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074648.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508074650.059, "dur": 32.384,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074650.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508074684.363, "dur": 1.152,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074684.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074690.347, "dur": 55.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074690.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074745.579, "dur": 48.320,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074745.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508074794.027, "dur": 2.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074794.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508074796.267, "dur": 20.448,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074796.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508074816.843, "dur": 6.592,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074816.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508074823.563, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074823.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074824.907, "dur": 13.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074824.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508074838.059, "dur": 2.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074838.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074841.099, "dur": 16.672,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074841.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508074857.899, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074857.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508074859.275, "dur": 9.312,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074859.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508074868.715, "dur": 3.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074868.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508074871.947, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074871.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508074873.131, "dur": 15.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074873.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508074888.299, "dur": 1.664,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074888.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074890.091, "dur": 13.440,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074890.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508074903.659, "dur": 2.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074903.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508074905.995, "dur": 3.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074905.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508074909.739, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074909.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508074910.827, "dur": 152.128,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508074910.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508075063.051, "dur": 6.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075063.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508075070.123, "dur": 4.768,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075070.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508075075.019, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075075.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508075076.651, "dur": 11.456,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075076.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508075088.235, "dur": 6.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075088.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508075095.307, "dur": 1.664,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075095.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508075097.099, "dur": 7.680,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075097.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508075104.939, "dur": 9.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075104.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508075114.123, "dur": 6.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075114.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075120.363, "dur": 22.496,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075120.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508075143.019, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075143.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075146.283, "dur": 26.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075146.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508075173.387, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075173.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508075174.987, "dur": 26.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075174.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508075201.163, "dur": 7.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075201.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508075208.331, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075208.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075209.675, "dur": 23.264,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075209.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508075233.035, "dur": 3.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075233.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075236.203, "dur": 29.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075236.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508075266.091, "dur": 1.632,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075266.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508075267.851, "dur": 31.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075267.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508075301.003, "dur": 0.992,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075301.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075306.795, "dur": 50.272,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075306.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075357.195, "dur": 47.968,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075357.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508075405.323, "dur": 2.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075405.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508075407.531, "dur": 22.464,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075407.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508075430.091, "dur": 6.816,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075430.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508075437.035, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075437.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075438.379, "dur": 13.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075438.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508075451.979, "dur": 2.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075451.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075454.923, "dur": 15.968,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075454.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508075471.019, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075471.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508075472.395, "dur": 9.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075472.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508075481.739, "dur": 3.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075481.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508075484.907, "dur": 1.088,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075484.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508075486.091, "dur": 15.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075486.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508075501.291, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075501.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075502.411, "dur": 13.440,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075502.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508075515.979, "dur": 2.176,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075515.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075518.283, "dur": 3.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075518.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508075522.027, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075522.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508075523.083, "dur": 153.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075523.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508075676.907, "dur": 7.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075676.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508075684.107, "dur": 4.832,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075684.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508075689.067, "dur": 1.472,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075689.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508075690.699, "dur": 11.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075690.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508075702.443, "dur": 6.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075702.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508075709.483, "dur": 1.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075709.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508075711.339, "dur": 8.384,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075711.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508075719.851, "dur": 10.656,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075719.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508075730.635, "dur": 6.016,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075730.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075736.811, "dur": 22.720,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075736.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508075759.659, "dur": 3.200,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075759.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075762.955, "dur": 27.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075762.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508075790.155, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075790.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508075791.787, "dur": 23.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075791.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508075814.955, "dur": 6.720,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075814.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508075821.803, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075821.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075823.147, "dur": 23.360,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075823.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508075846.635, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075846.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075849.899, "dur": 30.144,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075849.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508075880.171, "dur": 1.664,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075880.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508075881.931, "dur": 30.656,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075881.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508075915.243, "dur": 1.024,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075915.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075921.099, "dur": 52.128,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075921.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508075973.323, "dur": 48.512,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508075973.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508076021.963, "dur": 2.016,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076021.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508076024.107, "dur": 21.408,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076024.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508076045.643, "dur": 6.592,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076045.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508076052.363, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076052.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076053.707, "dur": 13.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076053.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508076067.051, "dur": 2.880,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076067.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076070.059, "dur": 16.832,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076070.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508076087.019, "dur": 1.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076087.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508076088.427, "dur": 9.184,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076088.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508076097.739, "dur": 3.328,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076097.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508076101.195, "dur": 1.056,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076101.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508076102.379, "dur": 15.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076102.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508076117.579, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076117.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076118.731, "dur": 13.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076118.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508076132.587, "dur": 2.272,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076132.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076134.987, "dur": 3.584,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076134.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508076138.699, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076138.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508076139.787, "dur": 159.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076139.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508076299.083, "dur": 6.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076299.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508076306.155, "dur": 4.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076306.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508076311.147, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076311.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508076312.843, "dur": 13.152,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076312.843,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508076326.123, "dur": 6.784,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076326.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508076333.035, "dur": 1.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076333.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508076334.859, "dur": 7.520,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076334.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508076342.475, "dur": 8.608,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076342.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508076351.179, "dur": 6.048,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076351.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076357.387, "dur": 22.176,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076357.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508076379.691, "dur": 3.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076379.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076383.019, "dur": 27.232,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076383.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508076410.379, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076410.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508076412.011, "dur": 23.424,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076412.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508076435.563, "dur": 6.784,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076435.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508076442.475, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076442.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076443.851, "dur": 23.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076443.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508076467.051, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076467.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076470.315, "dur": 29.792,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076470.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508076500.235, "dur": 1.600,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076500.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508076501.931, "dur": 32.192,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076501.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508076536.907, "dur": 1.088,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.1323529411764706
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076536.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076542.891, "dur": 51.200,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076542.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076594.219, "dur": 48.832,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076594.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508076643.179, "dur": 2.176,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076643.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508076645.483, "dur": 21.568,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076645.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508076667.147, "dur": 6.624,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076667.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508076673.899, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076673.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076675.275, "dur": 13.280,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076675.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508076688.683, "dur": 2.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076688.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076691.659, "dur": 16.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076691.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508076708.651, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076708.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508076709.995, "dur": 9.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076709.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508076719.851, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076719.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508076723.115, "dur": 1.088,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076723.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508076724.331, "dur": 15.520,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076724.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508076739.947, "dur": 1.568,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076739.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076741.675, "dur": 13.856,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076741.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508076755.691, "dur": 2.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076755.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076758.059, "dur": 3.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076758.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508076761.803, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076761.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508076762.891, "dur": 151.776,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076762.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508076914.795, "dur": 7.264,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076914.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508076922.155, "dur": 4.768,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076922.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508076927.019, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076927.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508076928.651, "dur": 11.872,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076928.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508076940.651, "dur": 6.816,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076940.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508076947.595, "dur": 1.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076947.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508076949.419, "dur": 8.768,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076949.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508076958.283, "dur": 10.400,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076958.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508076968.811, "dur": 5.888,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076968.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508076974.859, "dur": 23.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076974.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508076998.059, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508076998.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077001.323, "dur": 27.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077001.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508077028.619, "dur": 1.537,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077028.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508077030.283, "dur": 22.592,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077030.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508077052.971, "dur": 7.008,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077052.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508077060.107, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077060.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077061.483, "dur": 23.040,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077061.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508077084.619, "dur": 3.073,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077084.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077087.819, "dur": 30.144,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077087.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508077118.059, "dur": 1.632,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077118.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508077119.787, "dur": 31.168,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077119.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508077153.259, "dur": 1.024,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077153.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077159.083, "dur": 52.096,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077159.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077211.308, "dur": 48.543,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077211.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508077259.980, "dur": 2.047,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077259.980,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508077262.187, "dur": 21.088,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077262.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508077283.404, "dur": 6.623,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077283.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508077290.156, "dur": 1.247,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077290.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077291.499, "dur": 13.344,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077291.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508077304.972, "dur": 2.879,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077304.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077307.979, "dur": 16.736,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077307.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508077324.811, "dur": 1.249,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077324.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508077326.220, "dur": 8.959,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077326.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508077335.308, "dur": 3.071,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077335.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508077338.507, "dur": 1.088,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077338.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508077339.691, "dur": 15.776,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077339.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508077355.596, "dur": 1.023,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077355.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077356.747, "dur": 13.792,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077356.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508077370.667, "dur": 2.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077370.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077373.035, "dur": 3.680,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077373.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508077376.844, "dur": 0.927,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077376.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508077377.899, "dur": 156.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077377.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508077534.923, "dur": 7.041,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077534.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508077542.060, "dur": 4.800,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077542.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508077546.987, "dur": 1.505,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077546.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508077548.619, "dur": 12.224,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077548.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508077560.939, "dur": 6.689,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077560.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508077567.755, "dur": 1.729,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077567.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508077569.644, "dur": 7.679,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077569.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508077577.452, "dur": 9.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077577.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508077587.340, "dur": 6.175,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077587.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077593.644, "dur": 22.688,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077593.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508077616.459, "dur": 3.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077616.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077619.692, "dur": 27.007,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077619.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508077646.859, "dur": 1.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077646.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508077648.491, "dur": 24.384,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077648.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508077672.972, "dur": 6.944,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077672.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508077680.011, "dur": 1.249,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077680.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077681.387, "dur": 22.848,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077681.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508077704.364, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077704.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077707.596, "dur": 30.111,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077707.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508077737.836, "dur": 1.567,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077737.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508077739.532, "dur": 31.519,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077739.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508077773.804, "dur": 0.992,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077773.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077779.563, "dur": 51.297,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077779.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077830.987, "dur": 48.448,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077830.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508077879.564, "dur": 2.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077879.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508077881.804, "dur": 20.928,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077881.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508077902.859, "dur": 6.689,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077902.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508077909.675, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077909.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077911.020, "dur": 13.376,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077911.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508077924.524, "dur": 2.879,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077924.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077927.563, "dur": 16.929,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077927.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508077944.588, "dur": 1.247,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077944.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508077945.964, "dur": 9.792,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077945.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508077955.883, "dur": 3.041,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077955.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508077959.019, "dur": 1.089,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077959.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508077960.204, "dur": 14.880,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077960.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508077975.180, "dur": 2.015,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077975.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077977.324, "dur": 13.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077977.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508077991.211, "dur": 2.240,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077991.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508077993.580, "dur": 3.648,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077993.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508077997.356, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077997.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508077998.443, "dur": 157.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508077998.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508078156.140, "dur": 6.976,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078156.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508078163.244, "dur": 5.184,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078163.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508078168.556, "dur": 1.504,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078168.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508078170.187, "dur": 11.681,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078170.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508078181.964, "dur": 6.975,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078181.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508078189.068, "dur": 1.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078189.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508078190.924, "dur": 7.616,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078190.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508078198.636, "dur": 9.408,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078198.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508078208.171, "dur": 6.017,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078208.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078214.316, "dur": 23.232,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078214.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508078237.676, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078237.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078240.940, "dur": 27.232,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078240.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508078268.268, "dur": 1.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078268.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508078270.156, "dur": 22.560,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078270.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508078292.844, "dur": 6.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078292.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078299.884, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078299.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508078301.228, "dur": 19.039,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078301.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508078320.396, "dur": 13.440,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078320.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078333.964, "dur": 1.216,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078333.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078335.340, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078335.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508078336.716, "dur": 25.792,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078336.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078362.604, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078362.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078363.948, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078363.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508078365.324, "dur": 1.024,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078365.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508078366.508, "dur": 36.640,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078366.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508078403.276, "dur": 12.351,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078403.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078415.500, "dur": 1.888,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078415.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078417.516, "dur": 10.784,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078417.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508078428.428, "dur": 3.008,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078428.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078431.532, "dur": 11.552,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078431.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508078443.212, "dur": 1.152,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078443.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508078444.460, "dur": 7.488,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078444.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078452.044, "dur": 1.247,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078452.044,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508078453.420, "dur": 4.512,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078453.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508078458.060, "dur": 1.183,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078458.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078459.372, "dur": 10.176,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078459.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508078469.676, "dur": 2.816,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078469.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078472.620, "dur": 10.432,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078472.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508078483.212, "dur": 1.120,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078483.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290508078484.460, "dur": 2.528,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078484.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290508078487.148, "dur": 7.743,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078487.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078494.988, "dur": 22.784,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078494.988,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508078517.868, "dur": 3.072,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078517.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078521.068, "dur": 29.472,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078521.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508078550.668, "dur": 1.568,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078550.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508078552.364, "dur": 30.816,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078552.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508078585.996, "dur": 0.992,
    "args": {
      "External id": 5170,      "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078585.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078591.820, "dur": 53.696,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078591.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078645.644, "dur": 48.672,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078645.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508078694.444, "dur": 2.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078694.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290508078696.684, "dur": 33.408,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078696.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078730.220, "dur": 17.184,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078730.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508078747.532, "dur": 9.760,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078747.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508078757.420, "dur": 15.584,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078757.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078773.164, "dur": 5.856,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078773.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508078779.116, "dur": 159.584,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078779.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508078938.892, "dur": 13.408,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078938.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508078952.396, "dur": 7.424,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078952.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508078959.916, "dur": 6.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078959.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508078966.156, "dur": 4.864,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078966.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508078971.148, "dur": 5.920,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078971.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508078977.196, "dur": 27.872,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508078977.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290508079005.196, "dur": 20.384,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079005.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508079025.708, "dur": 1.248,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079025.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079027.084, "dur": 13.984,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079027.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508079041.196, "dur": 2.912,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079041.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508079044.236, "dur": 1.920,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079044.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508079046.284, "dur": 3.104,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079046.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508079049.516, "dur": 1.088,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079049.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508079050.732, "dur": 1.376,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079050.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079052.236, "dur": 13.664,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079052.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508079066.028, "dur": 2.112,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079066.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508079068.300, "dur": 0.960,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079068.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508079069.388, "dur": 4.896,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079069.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508079074.380, "dur": 1.568,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079074.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508079076.076, "dur": 7.392,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079076.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508079083.596, "dur": 1.728,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079083.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508079085.420, "dur": 22.080,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079085.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508079107.596, "dur": 3.136,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079107.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508079110.860, "dur": 1.536,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079110.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290508079112.492, "dur": 24.576,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079112.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508079137.196, "dur": 1.440,
    "args": {
      "External id": 5170,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 507906,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 507906, "pid": 0, "tid": 7, "ts": 1290508079137.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290508010207.634, "dur": 662.914,
    "args": {
      "External id": 5170,      "cbid": 311, "correlation": 507906
    }
  },
  {
    "ph": "s", "id": 507906, "pid": 14565, "tid": 14640, "ts": 1290508010207.634,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290508012964.969, "dur": 1.166,
    "args": {
      "External id": 5938,      "cbid": 317, "correlation": 508681
    }
  },
  {
    "ph": "f", "id": 508681, "pid": 14565, "tid": 14640, "ts": 1290508012964.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290508079141.356, "dur": 8.544,
    "args": {
      "External id": 5938,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508683,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508683, "pid": 0, "tid": 7, "ts": 1290508079141.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290508079150.252, "dur": 15.840,
    "args": {
      "External id": 5938,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508683,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508683, "pid": 0, "tid": 7, "ts": 1290508079150.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290508012966.976, "dur": 14.261,
    "args": {
      "External id": 5938,      "cbid": 311, "correlation": 508683
    }
  },
  {
    "ph": "s", "id": 508683, "pid": 14565, "tid": 14640, "ts": 1290508012966.976,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290508079167.244, "dur": 22.080,
    "args": {
      "External id": 5940,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508694,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508694, "pid": 0, "tid": 7, "ts": 1290508079167.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290508013022.481, "dur": 11.694,
    "args": {
      "External id": 5940,      "cbid": 211, "correlation": 508694
    }
  },
  {
    "ph": "s", "id": 508694, "pid": 14565, "tid": 14640, "ts": 1290508013022.481,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::multi_tensor_apply_kernel<at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float> >(at::native::(anonymous namespace)::TensorListMetadata<2>, at::native::(anonymous namespace)::UnaryOpFunctor<float, 2, 1, 1>, at::native::Copy<float, float>)", "pid": 0, "tid": 7,
    "ts": 1290508085632.749, "dur": 2.080,
    "args": {
      "External id": 5948,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508779,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508779, "pid": 0, "tid": 7, "ts": 1290508085632.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290508017100.460, "dur": 21.795,
    "args": {
      "External id": 5948,      "cbid": 211, "correlation": 508779
    }
  },
  {
    "ph": "s", "id": 508779, "pid": 14565, "tid": 14640, "ts": 1290508017100.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290508017131.970, "dur": 0.756,
    "args": {
      "External id": 5946,      "cbid": 317, "correlation": 508787
    }
  },
  {
    "ph": "f", "id": 508787, "pid": 14565, "tid": 14640, "ts": 1290508017131.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_div_mul_0", "pid": 0, "tid": 7,
    "ts": 1290508085638.989, "dur": 12.320,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085638.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_1", "pid": 0, "tid": 7,
    "ts": 1290508085651.437, "dur": 17.984,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085651.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_div_mul_view_2", "pid": 0, "tid": 7,
    "ts": 1290508085669.549, "dur": 6.560,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085669.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508085676.205, "dur": 1.344,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085676.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085677.677, "dur": 23.360,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085677.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508085701.165, "dur": 3.296,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085701.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085704.557, "dur": 30.272,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085704.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508085734.957, "dur": 1.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085734.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508085736.717, "dur": 31.872,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085736.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508085770.189, "dur": 1.024,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085770.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085776.109, "dur": 56.000,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085776.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085832.237, "dur": 48.800,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085832.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508085881.133, "dur": 2.080,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085881.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_7", "pid": 0, "tid": 7,
    "ts": 1290508085883.341, "dur": 29.696,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 128,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085883.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508085913.133, "dur": 6.592,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085913.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508085919.853, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085919.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085921.229, "dur": 13.408,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085921.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508085934.765, "dur": 2.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085934.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085937.805, "dur": 16.320,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085937.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508085954.285, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085954.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508085955.661, "dur": 9.472,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085955.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508085965.293, "dur": 3.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085965.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508085968.557, "dur": 1.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085968.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508085969.741, "dur": 15.392,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085969.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508085985.229, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085985.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508085986.381, "dur": 13.664,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508085986.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508086000.173, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086000.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086002.541, "dur": 3.680,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086002.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508086006.349, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086006.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508086007.437, "dur": 157.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086007.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508086164.685, "dur": 6.976,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086164.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508086171.821, "dur": 4.768,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086171.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508086176.717, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086176.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508086178.349, "dur": 11.840,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086178.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508086190.317, "dur": 6.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086190.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508086197.357, "dur": 1.696,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086197.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508086199.149, "dur": 7.488,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086199.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508086206.765, "dur": 8.608,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086206.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508086215.501, "dur": 5.760,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086215.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086221.389, "dur": 22.368,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086221.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508086243.885, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086243.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086247.117, "dur": 26.976,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086247.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508086274.221, "dur": 1.472,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086274.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508086275.853, "dur": 20.224,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086275.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508086296.237, "dur": 6.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086296.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508086303.149, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086303.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086304.493, "dur": 23.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086304.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508086327.725, "dur": 3.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086327.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086330.989, "dur": 29.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086330.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508086360.717, "dur": 1.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086360.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508086362.445, "dur": 31.872,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086362.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508086397.005, "dur": 1.024,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086397.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086402.893, "dur": 51.680,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086402.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086454.701, "dur": 48.448,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086454.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508086503.309, "dur": 2.048,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086503.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508086505.517, "dur": 21.376,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086505.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508086527.021, "dur": 6.624,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086527.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508086533.773, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086533.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086535.149, "dur": 13.344,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086535.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508086548.621, "dur": 2.816,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086548.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086551.597, "dur": 17.408,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086551.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508086569.165, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086569.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508086570.509, "dur": 9.664,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086570.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508086580.269, "dur": 2.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086580.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508086583.341, "dur": 1.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086583.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508086584.525, "dur": 15.328,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086584.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508086599.949, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086599.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086601.101, "dur": 13.888,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086601.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508086615.085, "dur": 2.272,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086615.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086617.453, "dur": 3.712,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086617.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508086621.293, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086621.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508086622.381, "dur": 164.160,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086622.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508086786.669, "dur": 7.360,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086786.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508086794.157, "dur": 4.768,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086794.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508086799.053, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086799.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508086800.685, "dur": 12.064,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086800.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508086812.877, "dur": 7.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086812.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508086820.045, "dur": 1.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086820.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508086821.869, "dur": 7.680,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086821.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508086829.677, "dur": 9.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086829.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508086839.565, "dur": 6.176,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086839.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086845.869, "dur": 22.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086845.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508086868.909, "dur": 3.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086868.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086872.109, "dur": 27.360,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086872.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508086899.597, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086899.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508086901.229, "dur": 25.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086901.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508086926.381, "dur": 6.816,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086926.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508086933.325, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086933.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086934.669, "dur": 23.296,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086934.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508086958.093, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086958.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508086961.325, "dur": 29.856,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086961.325,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508086991.309, "dur": 1.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086991.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508086993.037, "dur": 31.968,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508086993.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508087027.789, "dur": 1.025,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.1404878048780488
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087027.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087033.581, "dur": 50.688,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087033.581,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087084.397, "dur": 48.480,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087084.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508087133.037, "dur": 2.081,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087133.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508087135.278, "dur": 22.079,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087135.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508087157.485, "dur": 6.624,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087157.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508087164.269, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087164.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087165.614, "dur": 13.631,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087165.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508087179.373, "dur": 2.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087179.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087182.381, "dur": 16.672,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087182.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508087199.149, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087199.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508087200.493, "dur": 9.920,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087200.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508087210.509, "dur": 2.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087210.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508087213.613, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087213.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508087214.798, "dur": 15.263,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087214.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508087230.189, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087230.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087231.565, "dur": 14.112,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087231.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508087245.805, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087245.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087248.173, "dur": 3.680,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087248.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508087251.981, "dur": 0.961,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087251.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508087253.069, "dur": 171.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087253.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508087424.397, "dur": 7.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087424.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508087431.565, "dur": 4.864,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087431.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508087436.557, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087436.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508087438.221, "dur": 11.968,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087438.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508087450.317, "dur": 6.976,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087450.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508087457.389, "dur": 1.696,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087457.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508087459.213, "dur": 7.553,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087459.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508087466.861, "dur": 7.904,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087466.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508087474.894, "dur": 5.983,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087474.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087481.006, "dur": 22.623,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087481.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508087503.758, "dur": 3.167,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087503.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087506.990, "dur": 27.135,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087506.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508087534.253, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087534.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508087535.886, "dur": 25.119,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087535.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508087561.165, "dur": 6.913,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087561.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508087568.238, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087568.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087569.582, "dur": 23.295,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087569.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508087593.006, "dur": 3.039,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087593.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087596.173, "dur": 29.921,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087596.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508087626.221, "dur": 1.633,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087626.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508087627.950, "dur": 31.936,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087627.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508087662.638, "dur": 0.991,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.14530776992936428
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087662.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087668.397, "dur": 50.241,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087668.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087718.734, "dur": 48.479,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087718.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508087767.342, "dur": 2.080,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087767.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508087769.582, "dur": 21.663,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087769.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508087791.341, "dur": 6.625,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087791.341,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508087798.062, "dur": 1.247,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087798.062,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087799.405, "dur": 13.217,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087799.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508087812.749, "dur": 2.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087812.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087815.725, "dur": 16.065,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087815.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508087831.917, "dur": 1.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087831.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508087833.326, "dur": 9.343,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087833.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508087842.798, "dur": 3.199,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087842.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508087846.126, "dur": 1.055,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087846.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508087847.310, "dur": 14.783,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087847.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508087862.222, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087862.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087863.342, "dur": 13.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087863.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508087877.070, "dur": 2.239,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087877.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508087879.438, "dur": 3.680,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087879.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508087883.214, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087883.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508087884.301, "dur": 166.209,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508087884.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508088050.670, "dur": 6.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088050.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508088057.742, "dur": 4.896,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088057.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508088062.766, "dur": 1.535,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088062.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508088064.397, "dur": 11.713,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088064.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508088076.238, "dur": 6.975,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088076.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508088083.342, "dur": 1.696,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088083.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508088085.166, "dur": 8.384,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088085.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508088093.646, "dur": 9.984,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088093.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508088103.758, "dur": 6.048,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088103.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088109.902, "dur": 22.559,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088109.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508088132.621, "dur": 3.105,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088132.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088135.886, "dur": 27.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088135.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508088163.118, "dur": 1.535,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088163.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508088164.782, "dur": 23.712,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088164.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508088188.622, "dur": 6.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088188.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508088195.565, "dur": 1.249,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088195.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088196.942, "dur": 23.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088196.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508088220.174, "dur": 3.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088220.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088223.470, "dur": 29.631,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088223.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508088253.230, "dur": 1.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088253.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508088254.957, "dur": 31.425,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088254.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508088289.358, "dur": 1.024,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088289.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088295.278, "dur": 52.352,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088295.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088347.758, "dur": 48.448,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088347.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508088396.334, "dur": 2.048,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088396.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508088398.510, "dur": 20.736,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088398.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508088419.374, "dur": 6.656,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088419.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508088426.158, "dur": 1.247,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088426.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088427.502, "dur": 13.472,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088427.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508088441.070, "dur": 2.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088441.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088444.078, "dur": 16.896,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088444.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508088461.102, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088461.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508088462.446, "dur": 9.120,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088462.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508088471.694, "dur": 2.976,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088471.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508088474.798, "dur": 1.055,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088474.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508088475.982, "dur": 14.560,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088475.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508088490.670, "dur": 1.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088490.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088491.886, "dur": 13.888,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088491.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508088505.869, "dur": 2.369,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088505.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088508.366, "dur": 3.648,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088508.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508088512.142, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088512.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508088513.230, "dur": 158.528,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088513.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508088671.886, "dur": 6.976,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088671.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508088678.990, "dur": 4.800,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088678.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508088683.918, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088683.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508088685.550, "dur": 11.584,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088685.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508088697.262, "dur": 7.232,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088697.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508088704.622, "dur": 1.664,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088704.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508088706.446, "dur": 7.520,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088706.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508088714.094, "dur": 8.768,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088714.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508088722.958, "dur": 5.952,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088722.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088729.006, "dur": 22.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088729.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508088751.854, "dur": 3.200,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088751.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088755.150, "dur": 26.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088755.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508088782.158, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088782.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508088783.822, "dur": 28.416,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088783.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508088812.334, "dur": 6.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088812.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508088819.310, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088819.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088820.686, "dur": 23.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088820.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508088843.886, "dur": 3.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088843.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088847.054, "dur": 30.016,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088847.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508088877.230, "dur": 1.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088877.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508088878.958, "dur": 31.776,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088878.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508088912.814, "dur": 1.056,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088912.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088918.734, "dur": 51.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088918.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508088969.934, "dur": 48.544,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508088969.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508089018.606, "dur": 2.080,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089018.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508089020.814, "dur": 20.832,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089020.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508089041.742, "dur": 6.624,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089041.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508089048.462, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089048.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089049.838, "dur": 13.184,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089049.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508089063.150, "dur": 2.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089063.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089066.158, "dur": 16.800,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089066.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508089083.086, "dur": 1.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089083.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508089084.462, "dur": 8.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089084.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508089093.518, "dur": 3.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089093.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508089096.718, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089096.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508089097.902, "dur": 15.328,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089097.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508089113.326, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089113.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089114.478, "dur": 13.952,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089114.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508089128.558, "dur": 2.272,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089128.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089130.926, "dur": 3.648,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089130.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508089134.702, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089134.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508089135.790, "dur": 158.816,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089135.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508089294.766, "dur": 7.008,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089294.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508089301.902, "dur": 4.832,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089301.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508089306.894, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089306.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508089308.526, "dur": 12.000,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089308.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508089320.686, "dur": 7.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089320.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508089327.886, "dur": 1.696,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089327.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508089329.710, "dur": 7.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089329.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508089337.390, "dur": 8.672,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089337.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508089346.158, "dur": 5.888,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089346.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089352.142, "dur": 22.656,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089352.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508089374.926, "dur": 3.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089374.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089378.190, "dur": 27.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089378.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508089405.390, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089405.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508089406.990, "dur": 25.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089406.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508089432.846, "dur": 6.784,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089432.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508089439.758, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089439.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089441.134, "dur": 23.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089441.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508089464.366, "dur": 3.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089464.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089467.662, "dur": 29.920,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089467.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508089497.742, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089497.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508089499.406, "dur": 31.264,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089499.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508089533.486, "dur": 1.152,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089533.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089539.470, "dur": 49.984,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089539.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089589.582, "dur": 48.480,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089589.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508089638.190, "dur": 2.080,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089638.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508089640.366, "dur": 21.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089640.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508089662.190, "dur": 6.688,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089662.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508089668.974, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089668.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089670.350, "dur": 13.184,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089670.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508089683.694, "dur": 2.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089683.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089686.638, "dur": 16.576,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089686.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508089703.310, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089703.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508089704.686, "dur": 9.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089704.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508089713.934, "dur": 3.328,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089713.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508089717.422, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089717.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508089718.606, "dur": 15.008,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089718.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508089733.742, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089733.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089734.894, "dur": 13.824,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089734.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508089748.846, "dur": 2.208,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089748.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089751.214, "dur": 3.648,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089751.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508089754.990, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089754.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508089756.046, "dur": 159.200,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089756.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508089915.406, "dur": 7.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089915.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508089922.606, "dur": 4.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089922.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508089927.694, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089927.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508089929.326, "dur": 11.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089929.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508089941.006, "dur": 7.584,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089941.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508089948.686, "dur": 1.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089948.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508089950.510, "dur": 7.712,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089950.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508089958.350, "dur": 9.568,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089958.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508089968.014, "dur": 6.176,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089968.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508089974.286, "dur": 22.400,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089974.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508089996.814, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508089996.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090000.046, "dur": 27.264,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090000.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508090027.438, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090027.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508090029.070, "dur": 24.416,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090029.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508090053.614, "dur": 6.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090053.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090060.590, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090060.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090061.966, "dur": 23.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090061.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508090085.198, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090085.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090088.430, "dur": 30.144,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090088.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508090118.702, "dur": 1.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090118.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508090120.462, "dur": 31.840,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090120.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508090154.062, "dur": 1.056,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.13636363636363635
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090154.062,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090159.950, "dur": 52.736,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090159.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090212.814, "dur": 48.384,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090212.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508090261.326, "dur": 2.080,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090261.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508090263.534, "dur": 21.952,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090263.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508090285.582, "dur": 6.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090285.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090292.430, "dur": 1.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090292.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090293.806, "dur": 13.312,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090293.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508090307.246, "dur": 2.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090307.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090310.222, "dur": 16.608,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090310.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508090326.958, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090326.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508090328.334, "dur": 8.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090328.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508090337.422, "dur": 3.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090337.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508090340.622, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090340.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508090341.774, "dur": 15.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090341.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508090356.974, "dur": 1.408,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090356.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090358.510, "dur": 13.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090358.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508090372.142, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090372.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090374.510, "dur": 3.616,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090374.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508090378.254, "dur": 0.928,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090378.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508090379.310, "dur": 161.888,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090379.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508090541.326, "dur": 7.328,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090541.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508090548.782, "dur": 4.736,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090548.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508090553.646, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090553.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508090555.278, "dur": 11.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090555.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508090566.958, "dur": 6.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090566.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508090574.030, "dur": 1.664,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090574.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508090575.854, "dur": 7.648,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090575.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508090583.662, "dur": 9.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090583.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508090593.038, "dur": 5.952,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090593.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090599.118, "dur": 22.592,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090599.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508090621.806, "dur": 3.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090621.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090625.102, "dur": 26.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090625.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508090652.110, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090652.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508090653.742, "dur": 20.864,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090653.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508090674.734, "dur": 7.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090674.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090681.870, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090681.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508090683.214, "dur": 18.656,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090683.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508090701.966, "dur": 13.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090701.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090715.598, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090715.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090716.942, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090716.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508090718.318, "dur": 26.144,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090718.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090744.622, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090744.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090745.934, "dur": 1.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090745.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508090747.310, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090747.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508090748.494, "dur": 37.856,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090748.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508090786.510, "dur": 10.112,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090786.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090796.750, "dur": 1.856,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090796.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090798.734, "dur": 10.592,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090798.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508090809.454, "dur": 2.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090809.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090812.462, "dur": 11.808,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090812.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508090824.398, "dur": 1.152,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090824.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508090825.678, "dur": 7.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090825.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090833.358, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090833.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508090834.702, "dur": 4.416,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090834.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508090839.246, "dur": 1.184,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090839.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090840.590, "dur": 10.112,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090840.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508090850.830, "dur": 2.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090850.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090853.774, "dur": 10.592,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090853.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508090864.494, "dur": 1.152,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090864.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_39", "pid": 0, "tid": 7,
    "ts": 1290508090865.774, "dur": 2.496,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 5.595745,      "warps per SM": 22.382978,      "grid": [1052, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 47
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090865.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_40", "pid": 0, "tid": 7,
    "ts": 1290508090868.366, "dur": 7.392,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 46,      "shared memory": 8192,      "blocks per SM": 2.723404,      "warps per SM": 21.787233,      "grid": [1, 512, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 45
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090868.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090875.886, "dur": 22.816,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090875.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508090898.830, "dur": 3.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090898.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090902.030, "dur": 29.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090902.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508090931.790, "dur": 1.568,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090931.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508090933.486, "dur": 31.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090933.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508090965.966, "dur": 1.152,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.125
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090965.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508090971.918, "dur": 53.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508090971.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091025.326, "dur": 48.704,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091025.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508091074.158, "dur": 2.144,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091074.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508091076.430, "dur": 30.496,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091076.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508091107.054, "dur": 6.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091107.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508091114.062, "dur": 1.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091114.062,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091115.438, "dur": 13.344,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091115.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508091128.878, "dur": 2.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091128.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091131.886, "dur": 16.224,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091131.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508091148.270, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091148.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508091149.646, "dur": 9.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091149.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508091159.022, "dur": 2.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091159.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508091162.094, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091162.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508091163.278, "dur": 15.008,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091163.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508091178.414, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091178.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091179.566, "dur": 13.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091179.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508091193.230, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091193.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091195.566, "dur": 3.616,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091195.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508091199.310, "dur": 1.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091199.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508091200.526, "dur": 159.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091200.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508091359.886, "dur": 6.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091359.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508091366.958, "dur": 4.768,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091366.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508091371.854, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091371.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508091373.486, "dur": 12.416,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091373.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508091386.030, "dur": 6.784,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091386.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508091392.942, "dur": 1.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091392.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508091394.766, "dur": 7.520,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091394.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508091402.414, "dur": 9.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091402.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508091411.598, "dur": 5.952,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091411.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091417.678, "dur": 23.232,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091417.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508091441.038, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091441.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091444.270, "dur": 27.232,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091444.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508091471.630, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091471.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508091473.262, "dur": 23.328,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091473.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508091496.718, "dur": 6.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091496.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508091503.598, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091503.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091504.942, "dur": 22.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091504.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508091528.014, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091528.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091531.246, "dur": 29.856,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091531.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508091561.230, "dur": 1.696,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091561.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508091563.086, "dur": 32.288,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091563.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508091596.814, "dur": 0.992,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.14516129032258066
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091596.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091602.638, "dur": 50.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091602.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091653.518, "dur": 47.488,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091653.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508091701.134, "dur": 2.048,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091701.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508091703.310, "dur": 20.640,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091703.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508091724.078, "dur": 6.624,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091724.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508091730.830, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091730.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091732.174, "dur": 12.992,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091732.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508091745.294, "dur": 2.784,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091745.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091748.206, "dur": 16.576,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091748.206,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508091764.910, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091764.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508091766.286, "dur": 9.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091766.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508091775.694, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091775.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508091778.926, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091778.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508091780.110, "dur": 15.616,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091780.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508091795.854, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091795.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091797.486, "dur": 13.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091797.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508091811.246, "dur": 2.176,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091811.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508091813.550, "dur": 3.616,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091813.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508091817.294, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091817.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508091818.350, "dur": 154.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091818.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508091973.454, "dur": 6.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091973.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508091980.462, "dur": 4.833,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091980.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508091985.390, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091985.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508091986.990, "dur": 11.456,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091986.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508091998.575, "dur": 6.943,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508091998.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508092005.646, "dur": 1.697,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092005.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508092007.470, "dur": 7.681,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092007.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508092015.278, "dur": 9.120,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092015.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508092024.526, "dur": 6.016,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092024.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092030.670, "dur": 22.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092030.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508092053.038, "dur": 3.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092053.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092056.334, "dur": 26.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092056.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508092083.311, "dur": 1.567,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092083.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508092084.974, "dur": 25.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092084.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508092110.351, "dur": 6.943,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092110.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508092117.390, "dur": 1.249,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092117.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092118.734, "dur": 23.616,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092118.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508092142.446, "dur": 3.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092142.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092145.679, "dur": 29.759,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092145.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508092175.566, "dur": 1.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092175.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508092177.294, "dur": 31.073,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092177.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508092211.246, "dur": 0.993,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.14501510574018128
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092211.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092217.038, "dur": 51.744,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092217.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092268.910, "dur": 48.800,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092268.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508092317.838, "dur": 2.145,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092317.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508092320.110, "dur": 22.817,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092320.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508092343.054, "dur": 6.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092343.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508092349.934, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092349.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092351.279, "dur": 13.311,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092351.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508092364.719, "dur": 2.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092364.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092367.758, "dur": 16.129,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092367.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508092384.014, "dur": 1.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092384.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508092385.422, "dur": 9.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092385.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508092394.831, "dur": 3.231,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092394.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508092398.222, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092398.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508092399.407, "dur": 14.720,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092399.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508092414.223, "dur": 1.055,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092414.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092415.374, "dur": 13.729,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092415.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508092429.230, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092429.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092431.599, "dur": 3.647,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092431.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508092435.406, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092435.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508092436.462, "dur": 158.784,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092436.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508092595.342, "dur": 7.105,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092595.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508092602.574, "dur": 4.865,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092602.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508092607.566, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092607.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508092609.199, "dur": 11.584,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092609.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508092620.943, "dur": 6.911,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092620.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508092627.982, "dur": 1.729,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092627.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508092629.838, "dur": 8.064,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092629.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508092638.031, "dur": 10.432,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092638.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508092648.590, "dur": 5.952,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092648.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092654.671, "dur": 22.719,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092654.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508092677.519, "dur": 3.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092677.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092680.751, "dur": 27.360,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092680.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508092708.238, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092708.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508092709.871, "dur": 23.135,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092709.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508092733.102, "dur": 6.785,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092733.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508092739.983, "dur": 1.247,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092739.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092741.359, "dur": 23.328,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092741.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508092764.814, "dur": 3.073,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092764.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092768.014, "dur": 30.081,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092768.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508092798.223, "dur": 1.599,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092798.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508092799.950, "dur": 30.592,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092799.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508092833.807, "dur": 1.120,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092833.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092839.919, "dur": 52.864,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092839.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092892.910, "dur": 48.417,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092892.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508092941.454, "dur": 2.049,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092941.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508092943.631, "dur": 21.920,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092943.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508092965.679, "dur": 6.624,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092965.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508092972.430, "dur": 1.249,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092972.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092973.807, "dur": 13.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092973.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508092986.990, "dur": 2.849,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092986.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508092989.966, "dur": 16.801,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508092989.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508093006.895, "dur": 1.280,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093006.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508093008.271, "dur": 9.408,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093008.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508093017.806, "dur": 3.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093017.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508093021.038, "dur": 1.057,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093021.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508093022.223, "dur": 14.560,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093022.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508093036.911, "dur": 1.023,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093036.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093038.063, "dur": 13.695,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093038.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508093051.887, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093051.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093054.254, "dur": 3.649,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093054.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508093058.063, "dur": 0.959,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093058.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508093059.119, "dur": 162.687,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093059.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508093221.903, "dur": 6.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093221.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508093228.974, "dur": 4.865,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093228.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508093233.966, "dur": 1.537,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093233.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508093235.631, "dur": 12.735,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093235.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508093248.463, "dur": 6.751,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093248.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508093255.311, "dur": 1.792,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093255.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508093257.199, "dur": 7.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093257.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508093264.879, "dur": 8.576,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093264.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508093273.583, "dur": 5.888,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093273.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093279.599, "dur": 22.176,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093279.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508093301.871, "dur": 3.071,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093301.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093305.102, "dur": 27.553,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093305.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508093332.782, "dur": 1.505,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093332.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508093334.447, "dur": 23.488,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093334.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508093358.063, "dur": 6.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093358.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508093365.039, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093365.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093366.415, "dur": 22.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093366.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508093389.358, "dur": 3.105,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093389.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093392.591, "dur": 30.079,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093392.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508093422.799, "dur": 1.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093422.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508093424.495, "dur": 32.160,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093424.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508093459.503, "dur": 1.120,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.12857142857142856
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093459.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093465.423, "dur": 52.512,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093465.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093518.063, "dur": 48.480,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093518.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508093566.671, "dur": 2.112,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093566.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508093568.911, "dur": 21.568,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093568.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508093590.575, "dur": 6.656,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093590.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508093597.327, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093597.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093598.703, "dur": 13.184,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093598.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508093612.015, "dur": 2.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093612.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093614.991, "dur": 17.440,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093614.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508093632.527, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093632.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508093633.903, "dur": 9.568,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093633.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508093643.599, "dur": 3.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093643.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508093646.863, "dur": 1.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093646.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508093648.079, "dur": 15.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093648.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508093663.759, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093663.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093664.911, "dur": 13.888,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093664.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508093678.927, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093678.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093681.295, "dur": 3.648,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093681.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508093685.071, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093685.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508093686.159, "dur": 163.648,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093686.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508093849.935, "dur": 7.392,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093849.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508093857.455, "dur": 4.768,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093857.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508093862.319, "dur": 1.568,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093862.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508093863.983, "dur": 11.808,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093863.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508093875.919, "dur": 6.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093875.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508093882.799, "dur": 1.664,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093882.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508093884.591, "dur": 8.512,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093884.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508093893.231, "dur": 10.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093893.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508093904.111, "dur": 6.368,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093904.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093910.639, "dur": 22.400,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093910.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508093933.167, "dur": 3.168,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093933.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093936.463, "dur": 26.816,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093936.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508093963.407, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093963.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508093965.039, "dur": 22.528,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093965.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508093987.663, "dur": 6.944,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093987.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508093994.703, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093994.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508093996.079, "dur": 23.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508093996.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508094019.343, "dur": 3.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094019.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094022.511, "dur": 29.824,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094022.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508094052.431, "dur": 1.568,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094052.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508094054.127, "dur": 30.720,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094054.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508094086.127, "dur": 1.024,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.140625
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094086.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094091.983, "dur": 49.856,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094091.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094141.935, "dur": 47.968,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094141.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508094190.031, "dur": 2.016,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094190.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508094192.207, "dur": 21.152,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094192.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508094213.487, "dur": 6.624,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094213.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508094220.239, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094220.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094221.583, "dur": 13.120,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094221.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508094234.831, "dur": 2.848,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094234.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094237.807, "dur": 16.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094237.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508094254.895, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094254.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508094256.239, "dur": 9.344,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094256.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508094265.679, "dur": 3.296,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094265.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508094269.103, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094269.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508094270.287, "dur": 15.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094270.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508094285.519, "dur": 0.992,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094285.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094286.639, "dur": 13.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094286.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508094300.399, "dur": 2.208,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094300.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094302.703, "dur": 3.616,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094302.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508094306.447, "dur": 0.928,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094306.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508094307.503, "dur": 158.080,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094307.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508094465.743, "dur": 7.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094465.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508094472.911, "dur": 4.864,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094472.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508094477.871, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094477.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508094479.503, "dur": 11.904,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094479.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508094491.535, "dur": 6.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094491.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508094498.447, "dur": 1.664,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094498.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508094500.271, "dur": 7.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094500.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508094507.951, "dur": 9.760,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094507.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508094517.871, "dur": 6.144,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094517.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094524.143, "dur": 23.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094524.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508094547.375, "dur": 3.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094547.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094550.639, "dur": 27.296,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094550.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508094578.063, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094578.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_26", "pid": 0, "tid": 7,
    "ts": 1290508094579.727, "dur": 24.320,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094579.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508094604.143, "dur": 6.688,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094604.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508094610.959, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094610.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094612.335, "dur": 22.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094612.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508094635.343, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094635.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094638.543, "dur": 29.856,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094638.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508094668.527, "dur": 1.568,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094668.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508094670.223, "dur": 30.912,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094670.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508094701.839, "dur": 0.480,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094701.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094707.215, "dur": 52.512,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094707.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094759.855, "dur": 48.448,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094759.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508094808.431, "dur": 2.080,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094808.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_view_28", "pid": 0, "tid": 7,
    "ts": 1290508094810.639, "dur": 21.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 90,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094810.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_view_8", "pid": 0, "tid": 7,
    "ts": 1290508094831.855, "dur": 6.656,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094831.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508094838.639, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094838.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094839.983, "dur": 13.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094839.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508094853.647, "dur": 2.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094853.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094856.655, "dur": 16.768,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094856.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508094873.551, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094873.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508094874.895, "dur": 9.120,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094874.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508094884.143, "dur": 3.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094884.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508094887.279, "dur": 1.088,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094887.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508094888.463, "dur": 15.424,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094888.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508094903.983, "dur": 1.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094903.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094905.775, "dur": 13.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094905.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508094919.631, "dur": 2.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094919.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508094921.999, "dur": 3.584,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094921.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508094925.743, "dur": 0.928,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094925.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508094926.799, "dur": 159.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508094926.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508095086.031, "dur": 7.008,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095086.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508095093.167, "dur": 5.120,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095093.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508095098.447, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095098.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508095100.079, "dur": 11.648,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095100.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508095111.823, "dur": 7.008,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095111.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508095118.927, "dur": 1.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095118.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508095120.751, "dur": 7.552,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095120.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508095128.399, "dur": 9.376,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095128.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508095137.903, "dur": 6.048,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095137.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095144.079, "dur": 22.688,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095144.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508095166.927, "dur": 3.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095166.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095170.159, "dur": 27.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095170.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508095197.391, "dur": 1.504,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095197.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_29", "pid": 0, "tid": 7,
    "ts": 1290508095199.055, "dur": 23.104,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095199.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward__to_copy_add_slice_backward_view_27", "pid": 0, "tid": 7,
    "ts": 1290508095222.287, "dur": 7.008,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 38,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095222.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095229.423, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095229.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_addcmul_mul_select_slice_30", "pid": 0, "tid": 7,
    "ts": 1290508095230.799, "dur": 19.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 16.340425,      "warps per SM": 65.361702,      "grid": [3072, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095230.799,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused_mul_silu_silu_backward_slice_sum_31", "pid": 0, "tid": 7,
    "ts": 1290508095249.935, "dur": 13.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 56,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095249.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095263.631, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095263.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095265.007, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095265.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_mul_select_silu_silu_backward_slice_backward_32", "pid": 0, "tid": 7,
    "ts": 1290508095266.383, "dur": 25.408,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 56,      "shared memory": 0,      "blocks per SM": 16.484043,      "warps per SM": 65.936172,      "grid": [3099, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 75
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095266.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095291.887, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095291.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095293.231, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095293.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_add_select_backward_view_33", "pid": 0, "tid": 7,
    "ts": 1290508095294.575, "dur": 1.024,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 0,      "blocks per SM": 0.031915,      "warps per SM": 0.127660,      "grid": [6, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095294.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_abs_add_clamp_min_constant_pad_nd_div_expand_ge_logsumexp_mul_sigmoid_sigmoid_backward_sign_sqrt_squeeze_sum_unsqueeze_view_where_34", "pid": 0, "tid": 7,
    "ts": 1290508095295.759, "dur": 37.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 254,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095295.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__unsafe_view_abs_clamp_min_constant_pad_nd_mul_sigmoid_sign_sqrt_unsqueeze_35", "pid": 0, "tid": 7,
    "ts": 1290508095333.135, "dur": 12.800,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 56,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095333.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095345.903, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095345.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095347.279, "dur": 10.816,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095347.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508095358.191, "dur": 2.976,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095358.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095361.295, "dur": 12.032,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095361.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508095373.455, "dur": 1.184,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095373.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_37", "pid": 0, "tid": 7,
    "ts": 1290508095374.735, "dur": 7.584,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095374.735,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095382.415, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095382.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm_backward_38", "pid": 0, "tid": 7,
    "ts": 1290508095383.791, "dur": 4.480,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 26,      "shared memory": 256,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095383.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095388.399, "dur": 1.216,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095388.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095389.743, "dur": 10.240,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.893617,      "warps per SM": 3.574468,      "grid": [6, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095389.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508095400.111, "dur": 2.816,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.021277,      "warps per SM": 16.340425,      "grid": [8, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095400.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095403.087, "dur": 10.432,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 0.680851,      "warps per SM": 2.723404,      "grid": [128, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095403.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_36", "pid": 0, "tid": 7,
    "ts": 1290508095413.679, "dur": 1.152,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.510638,      "warps per SM": 2.042553,      "grid": [96, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 4
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095413.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_41", "pid": 0, "tid": 7,
    "ts": 1290508095414.927, "dur": 2.624,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 10.989362,      "warps per SM": 87.914894,      "grid": [2066, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095414.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_add_embedding_dense_backward_logsumexp_view_42", "pid": 0, "tid": 7,
    "ts": 1290508095417.679, "dur": 7.680,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 32,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095417.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095425.487, "dur": 23.136,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095425.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508095448.751, "dur": 3.040,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [24, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095448.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095451.919, "dur": 29.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 232,      "shared memory": 65536,      "blocks per SM": 2.723404,      "warps per SM": 10.893617,      "grid": [512, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095451.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508095481.647, "dur": 1.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095481.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__unsafe_view_cat_mul_silu_silu_backward_split_view_5", "pid": 0, "tid": 7,
    "ts": 1290508095483.343, "dur": 32.512,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 65.361702,      "warps per SM": 261.446808,      "grid": [12288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095483.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memset", "name": "Memset (Device)", "pid": 0, "tid": 7,
    "ts": 1290508095517.999, "dur": 0.480,
    "args": {
      "External id": 5946,      "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "bytes": 144, "memory bandwidth (GB/s)": 0.3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095517.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095523.343, "dur": 54.752,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [48, 1, 4],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095523.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095578.223, "dur": 48.640,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095578.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_6", "pid": 0, "tid": 7,
    "ts": 1290508095626.991, "dur": 2.144,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 3.063830,      "warps per SM": 12.255320,      "grid": [576, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 26
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095626.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_view_43", "pid": 0, "tid": 7,
    "ts": 1290508095629.263, "dur": 32.704,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 96,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 42
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095629.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095662.095, "dur": 17.152,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095662.095,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_mul_neg_sigmoid_slice_sum_tanh_transpose_unsqueeze_view_10", "pid": 0, "tid": 7,
    "ts": 1290508095679.375, "dur": 9.696,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 58,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095679.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy__unsafe_view_add_clone_div_expand_flex_attention_backward_mul_neg_sigmoid_sigmoid_backward_slice_squeeze_sum_tanh_transpose_unsqueeze_view_13", "pid": 0, "tid": 7,
    "ts": 1290508095689.199, "dur": 17.376,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 128,      "blocks per SM": 16.340425,      "warps per SM": 32.680851,      "grid": [3072, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095689.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2>(cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_nn_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095706.703, "dur": 4.992,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 86,      "shared memory": 37376,      "blocks per SM": 1.361702,      "warps per SM": 5.446808,      "grid": [256, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 11
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095706.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_tem_fused__unsafe_view_flex_attention_backward_mul_sigmoid_transpose_unsqueeze_16", "pid": 0, "tid": 7,
    "ts": 1290508095711.823, "dur": 158.656,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 255,      "shared memory": 25088,      "blocks per SM": 6.127660,      "warps per SM": 24.510639,      "grid": [24, 8, 6],      "block": [128, 1, 1],      "est. achieved occupancy %": 17
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095711.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_20", "pid": 0, "tid": 7,
    "ts": 1290508095870.639, "dur": 13.632,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 65.361702,      "warps per SM": 130.723404,      "grid": [12288, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095870.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_23", "pid": 0, "tid": 7,
    "ts": 1290508095884.399, "dur": 7.360,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 76,      "shared memory": 16384,      "blocks per SM": 4.085106,      "warps per SM": 32.680851,      "grid": [6, 128, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095884.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_cat_mul_neg_slice_unsqueeze_17", "pid": 0, "tid": 7,
    "ts": 1290508095891.887, "dur": 6.112,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 35,      "shared memory": 32,      "blocks per SM": 32.680851,      "warps per SM": 65.361702,      "grid": [6144, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095891.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_24", "pid": 0, "tid": 7,
    "ts": 1290508095898.127, "dur": 4.928,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 4096,      "blocks per SM": 10.893617,      "warps per SM": 43.574467,      "grid": [1, 2048, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 91
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095898.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_cat_clone_sum_transpose_view_25", "pid": 0, "tid": 7,
    "ts": 1290508095903.151, "dur": 5.792,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095903.151,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3_nn_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095909.071, "dur": 28.000,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 73728,      "blocks per SM": 0.680851,      "warps per SM": 5.446808,      "grid": [128, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095909.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward__to_copy__unsafe_view_add_slice_backward_view_44", "pid": 0, "tid": 7,
    "ts": 1290508095937.231, "dur": 20.448,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 80,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 50
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095937.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508095957.807, "dur": 1.248,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095957.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095959.183, "dur": 13.984,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 0.957447,      "warps per SM": 3.829787,      "grid": [12, 1, 15],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095959.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508095973.295, "dur": 2.880,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 1.531915,      "warps per SM": 24.510639,      "grid": [12, 24, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 51
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095973.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_9", "pid": 0, "tid": 7,
    "ts": 1290508095976.303, "dur": 2.144,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 0.765957,      "warps per SM": 3.063830,      "grid": [144, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 6
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095976.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__to_copy_mul_sum_11", "pid": 0, "tid": 7,
    "ts": 1290508095978.575, "dur": 3.424,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 47,      "shared memory": 1024,      "blocks per SM": 0.680851,      "warps per SM": 1.361702,      "grid": [64, 2, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095978.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__to_copy_mul_sum_tanh_tanh_backward_view_12", "pid": 0, "tid": 7,
    "ts": 1290508095982.159, "dur": 1.056,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 32,      "blocks per SM": 0.010638,      "warps per SM": 0.021277,      "grid": [2, 1, 1],      "block": [64, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095982.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy__unsafe_view_clone_mm_mul_sigmoid_sigmoid_backward_squeeze_t_transpose_14", "pid": 0, "tid": 7,
    "ts": 1290508095983.343, "dur": 1.440,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 1.361702,      "warps per SM": 10.893617,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 23
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095983.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2>(cutlass_80_tensorop_bf16_s16816gemm_bf16_256x64_32x4_nt_align2::Params)", "pid": 0, "tid": 7,
    "ts": 1290508095984.911, "dur": 13.920,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 81920,      "blocks per SM": 0.074468,      "warps per SM": 0.297872,      "grid": [1, 1, 14],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095984.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508095998.959, "dur": 2.112,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.085106,      "grid": [1, 1, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508095998.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_mm_15", "pid": 0, "tid": 7,
    "ts": 1290508096001.199, "dur": 0.960,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.005319,      "warps per SM": 0.021277,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096001.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_18", "pid": 0, "tid": 7,
    "ts": 1290508096002.287, "dur": 4.896,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 4.085106,      "warps per SM": 16.340425,      "grid": [768, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 34
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096002.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_19", "pid": 0, "tid": 7,
    "ts": 1290508096007.279, "dur": 1.600,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 16,      "blocks per SM": 0.170213,      "warps per SM": 0.680851,      "grid": [32, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 1
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096007.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_21", "pid": 0, "tid": 7,
    "ts": 1290508096008.975, "dur": 7.456,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 40,      "shared memory": 2048,      "blocks per SM": 8.170213,      "warps per SM": 32.680851,      "grid": [1536, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 68
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096008.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_red_fused__fused_rms_norm__fused_rms_norm_backward_22", "pid": 0, "tid": 7,
    "ts": 1290508096016.559, "dur": 1.728,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 16,      "shared memory": 32,      "blocks per SM": 0.170213,      "warps per SM": 1.361702,      "grid": [32, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 3
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096016.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8>(cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x128_32x4_nt_align8::Params)", "pid": 0, "tid": 7,
    "ts": 1290508096018.383, "dur": 22.304,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 228,      "shared memory": 65536,      "blocks per SM": 1.021277,      "warps per SM": 4.085106,      "grid": [24, 1, 8],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096018.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cublasLt::splitKreduce_kernel<32, 16, int, __nv_bfloat16, __nv_bfloat16, float, __nv_bfloat16, false, __nv_bfloat16, __nv_bfloat16, __nv_bfloat16, true, false, false, false>(cublasLt::cublasSplitKParams<float>, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, __nv_bfloat16*, float const*, float const*, __nv_bfloat16 const*, __nv_bfloat16 const*, __nv_bfloat16*, void*, long, float*, int*, float*, float*, float const*, float const*, float const*, float const*, float const*)", "pid": 0, "tid": 7,
    "ts": 1290508096040.783, "dur": 3.072,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 42,      "shared memory": 0,      "blocks per SM": 3.063830,      "warps per SM": 49.021278,      "grid": [12, 48, 1],      "block": [32, 16, 1],      "est. achieved occupancy %": 67
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096040.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused__to_copy_4", "pid": 0, "tid": 7,
    "ts": 1290508096043.983, "dur": 1.536,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 20,      "shared memory": 2048,      "blocks per SM": 1.531915,      "warps per SM": 6.127660,      "grid": [288, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096043.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm__fused_rms_norm_backward__to_copy_add_slice_backward_view_45", "pid": 0, "tid": 7,
    "ts": 1290508096045.647, "dur": 24.704,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 106,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 21.787233,      "grid": [4096, 1, 1],      "block": [32, 1, 1],      "est. achieved occupancy %": 33
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096045.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_per_fused__fused_rms_norm_backward__to_copy_div_mul_view_3", "pid": 0, "tid": 7,
    "ts": 1290508096070.447, "dur": 1.472,
    "args": {
      "External id": 5946,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 508789,      "registers per thread": 22,      "shared memory": 128,      "blocks per SM": 0.255319,      "warps per SM": 1.021277,      "grid": [48, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 2
    }
  },
  {
    "ph": "f", "id": 508789, "pid": 0, "tid": 7, "ts": 1290508096070.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290508017133.593, "dur": 1328.200,
    "args": {
      "External id": 5946,      "cbid": 311, "correlation": 508789
    }
  },
  {
    "ph": "s", "id": 508789, "pid": 14565, "tid": 14640, "ts": 1290508017133.593,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 14565, "tid": 14640,
    "ts": 1290508021291.853, "dur": 0.683,
    "args": {
      "External id": 6714,      "cbid": 317, "correlation": 509564
    }
  },
  {
    "ph": "f", "id": 509564, "pid": 14565, "tid": 14640, "ts": 1290508021291.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 0, "tid": 7,
    "ts": 1290508096075.247, "dur": 8.544,
    "args": {
      "External id": 6714,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 509566,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 49.021278,      "warps per SM": 392.170227,      "grid": [9216, 1, 1],      "block": [256, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 509566, "pid": 0, "tid": 7, "ts": 1290508096075.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 0, "tid": 7,
    "ts": 1290508096084.143, "dur": 15.936,
    "args": {
      "External id": 6714,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 509566,      "registers per thread": 30,      "shared memory": 0,      "blocks per SM": 21.787233,      "warps per SM": 87.148933,      "grid": [1, 4096, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 509566, "pid": 0, "tid": 7, "ts": 1290508096084.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGraphLaunch", "pid": 14565, "tid": 14640,
    "ts": 1290508021293.314, "dur": 14.105,
    "args": {
      "External id": 6714,      "cbid": 311, "correlation": 509566
    }
  },
  {
    "ph": "s", "id": 509566, "pid": 14565, "tid": 14640, "ts": 1290508021293.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 0, "tid": 7,
    "ts": 1290508096101.135, "dur": 22.176,
    "args": {
      "External id": 6716,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 509577,      "registers per thread": 40,      "shared memory": 0,      "blocks per SM": 24.510639,      "warps per SM": 98.042557,      "grid": [4608, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 100
    }
  },
  {
    "ph": "f", "id": 509577, "pid": 0, "tid": 7, "ts": 1290508096101.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 14565, "tid": 14640,
    "ts": 1290508021344.812, "dur": 9.828,
    "args": {
      "External id": 6716,      "cbid": 211, "correlation": 509577
    }
  },
  {
    "ph": "s", "id": 509577, "pid": 14565, "tid": 14640, "ts": 1290508021344.812,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "overhead", "name": "Buffer Flush", "pid": -1, "tid": 0,
    "ts": 1290508003045.592, "dur": 194.015
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290508096075.246, "dur": 24.834,
    "args": {
      "External id": 6715
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290508062244.328, "dur": 24.897,
    "args": {
      "External id": 5163
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290508051838.693, "dur": 10401.988,
    "args": {
      "External id": 4395
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290508045353.444, "dur": 24.610,
    "args": {
      "External id": 4387
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290508034948.866, "dur": 10402.052,
    "args": {
      "External id": 3619
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290508028477.920, "dur": 24.931,
    "args": {
      "External id": 3611
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290508018073.375, "dur": 10401.220,
    "args": {
      "External id": 2843
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290508028528.641, "dur": 30.082,
    "args": {
      "External id": 127
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#6", "pid": 0, "tid": 7,
    "ts": 1290508018071.359, "dur": 1.570,
    "args": {
      "External id": 91
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290508011654.973, "dur": 6414.468,
    "args": {
      "External id": 104
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290508068732.777, "dur": 10405.860,
    "args": {
      "External id": 5171
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290508011624.509, "dur": 28.930,
    "args": {
      "External id": 97
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290507994760.954, "dur": 6406.499,
    "args": {
      "External id": 74
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290507994728.474, "dur": 29.666,
    "args": {
      "External id": 67
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#4", "pid": 0, "tid": 7,
    "ts": 1290507984271.160, "dur": 1.506,
    "args": {
      "External id": 31
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290507984273.112, "dur": 10402.820,
    "args": {
      "External id": 1291
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290507977840.726, "dur": 29.603,
    "args": {
      "External id": 37
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#10", "pid": 0, "tid": 7,
    "ts": 1290508085630.796, "dur": 1.538,
    "args": {
      "External id": 211
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#5", "pid": 0, "tid": 7,
    "ts": 1290508001169.211, "dur": 1.538,
    "args": {
      "External id": 61
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290508001171.195, "dur": 10401.540,
    "args": {
      "External id": 2067
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#3", "pid": 0, "tid": 7,
    "ts": 1290507967379.252, "dur": 1.570,
    "args": {
      "External id": 1
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290507960991.507, "dur": 6385.059,
    "args": {
      "External id": 14
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290508079189.739, "dur": 29.762,
    "args": {
      "External id": 217
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290507960958.995, "dur": 30.050,
    "args": {
      "External id": 7
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#9", "pid": 0, "tid": 7,
    "ts": 1290508068730.697, "dur": 1.666,
    "args": {
      "External id": 181
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290507977873.143, "dur": 6395.459,
    "args": {
      "External id": 44
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290507977791.798, "dur": 24.898,
    "args": {
      "External id": 1283
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290508028561.153, "dur": 6382.851,
    "args": {
      "External id": 134
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290508011575.773, "dur": 24.866,
    "args": {
      "External id": 2835
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#7", "pid": 0, "tid": 7,
    "ts": 1290508034946.882, "dur": 1.538,
    "args": {
      "External id": 121
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290508079141.355, "dur": 24.738,
    "args": {
      "External id": 5939
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290508045401.924, "dur": 30.594,
    "args": {
      "External id": 157
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290508045434.628, "dur": 6400.195,
    "args": {
      "External id": 164
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "ProfilerStep#8", "pid": 0, "tid": 7,
    "ts": 1290508051836.741, "dur": 1.506,
    "args": {
      "External id": 151
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fkft7abqdub5fbzrxsusa6o556tatvops7azzsrctjlvuoznpv4i ##", "pid": 0, "tid": 7,
    "ts": 1290508062292.807, "dur": 30.274,
    "args": {
      "External id": 187
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290508062325.511, "dur": 6402.148,
    "args": {
      "External id": 194
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290508085632.748, "dur": 10439.172,
    "args": {
      "External id": 5947
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph f2papzyhk3eoy65khzkyflsjbzrxv3wwrht5ce2klzmwjmzvt2jd ##", "pid": 0, "tid": 7,
    "ts": 1290508079220.619, "dur": 6408.835,
    "args": {
      "External id": 224
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph felvw6fyfphsgtviq7567hgcflv56pwuwthnvcc7ewpn5uqkdi3s ##", "pid": 0, "tid": 7,
    "ts": 1290507967381.237, "dur": 10408.323,
    "args": {
      "External id": 515
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "## Call CompiledFxGraph fcojz2ouyxh6vawborf6qurxv5ed25nb3zrbj5cfrmndj3wa6mav ##", "pid": 0, "tid": 7,
    "ts": 1290507994677.690, "dur": 24.994,
    "args": {
      "External id": 2059
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 0,
    "args": {
      "sort_index": 14565
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 0, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 1, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 2, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 3, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 4, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 5, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 6, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 7, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 8, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 9, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 10, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 11, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 12, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 13, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 14, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1290507954056.722, "pid": 15, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1290507954056.722, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1290507954056.722, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 14640,
    "args": {
      "name": "thread 14640 (python3)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 14640,
    "args": {
      "sort_index": 14640
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 14640,
    "args": {
      "name": "thread 14640 (pt_autograd_0)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 14640,
    "args": {
      "sort_index": 14640
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 14565,
    "args": {
      "name": "thread 14565 (python3)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1290507954056.722, "pid": 14565, "tid": 14565,
    "args": {
      "sort_index": 14565
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 1290507954022.743, "dur": 142108.478,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1290507954022.743,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 1290507954022.743
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 1290508109022.797
  }
  ],
  "traceName": "./profile_trace.json"
}