{
    "schemaVersion": 1,
    "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 1, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 2, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 3, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 4, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 5, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 6, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 7, "name": "NVIDIA H100 80GB HBM3", "totalGlobalMem": 85028896768,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    }
    ],
      "trace_id": "66FC132CED374E509323BAC9A429FB68",
      "cuda_driver_version": 12080,
      "cuda_runtime_version": 12080,
      "cupti_version": 26,
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1782967788000000000,
  "traceEvents": [
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#2", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400331.274, "dur": 244.565,
    "args": {
      "External id": 1,"Record function id": 0, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "flash_kernel_fwd", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400341.524, "dur": 228.975,
    "args": {
      "External id": 2,"Record function id": 0, "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400360.535, "dur": 200.034,
    "args": {
      "External id": 3,"Record function id": 0, "Sequence number": 5, "Fwd thread id": 0, "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn3_cuda_8a730d9::_flash_attn_forward", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400377.355, "dur": 173.653,
    "args": {
      "External id": 4,"Record function id": 0, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn3_cuda_8a730d9::fwd", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400412.476, "dur": 78.011,
    "args": {
      "External id": 5,"Record function id": 0, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400424.756, "dur": 14.050,
    "args": {
      "External id": 6,"Record function id": 0, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400430.606, "dur": 6.570,
    "args": {
      "External id": 7,"Record function id": 0, "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400440.836, "dur": 2.680,
    "args": {
      "External id": 8,"Record function id": 0, "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400441.096, "dur": 2.230,
    "args": {
      "External id": 9,"Record function id": 0, "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400445.126, "dur": 2.230,
    "args": {
      "External id": 10,"Record function id": 0, "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400445.366, "dur": 1.850,
    "args": {
      "External id": 11,"Record function id": 0, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400449.087, "dur": 26.080,
    "args": {
      "External id": 12,"Record function id": 0, "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400455.797, "dur": 18.360,
    "args": {
      "External id": 13,"Record function id": 0, "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400498.177, "dur": 1.130,
    "args": {
      "External id": 14,"Record function id": 0, "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400507.238, "dur": 15.680,
    "args": {
      "External id": 15,"Record function id": 0, "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400508.178, "dur": 14.600,
    "args": {
      "External id": 16,"Record function id": 0, "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400512.208, "dur": 2.500,
    "args": {
      "External id": 17,"Record function id": 0, "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400520.188, "dur": 2.330,
    "args": {
      "External id": 18,"Record function id": 0, "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400526.068, "dur": 0.170,
    "args": {
      "External id": 19,"Record function id": 0, "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400527.288, "dur": 0.140,
    "args": {
      "External id": 20,"Record function id": 0, "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400532.758, "dur": 0.270,
    "args": {
      "External id": 21,"Record function id": 0, "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400533.528, "dur": 4.150,
    "args": {
      "External id": 22,"Record function id": 0, "Ev Idx": 21
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400533.768, "dur": 3.770,
    "args": {
      "External id": 23,"Record function id": 0, "Ev Idx": 22
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400534.368, "dur": 2.060,
    "args": {
      "External id": 24,"Record function id": 0, "Ev Idx": 23
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400536.788, "dur": 0.600,
    "args": {
      "External id": 25,"Record function id": 0, "Ev Idx": 24
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400538.198, "dur": 0.050,
    "args": {
      "External id": 26,"Record function id": 0, "Ev Idx": 25
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400538.608, "dur": 0.080,
    "args": {
      "External id": 27,"Record function id": 0, "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#3", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400588.139, "dur": 170.193,
    "args": {
      "External id": 28,"Record function id": 0, "Ev Idx": 27
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "flash_kernel_fwd", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400595.739, "dur": 157.793,
    "args": {
      "External id": 29,"Record function id": 0, "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400609.920, "dur": 135.232,
    "args": {
      "External id": 30,"Record function id": 0, "Sequence number": 6, "Fwd thread id": 0, "Ev Idx": 29
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn3_cuda_8a730d9::_flash_attn_forward", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400624.010, "dur": 112.432,
    "args": {
      "External id": 31,"Record function id": 0, "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn3_cuda_8a730d9::fwd", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400650.650, "dur": 49.691,
    "args": {
      "External id": 32,"Record function id": 0, "Ev Idx": 31
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400655.250, "dur": 5.691,
    "args": {
      "External id": 33,"Record function id": 0, "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400655.880, "dur": 4.741,
    "args": {
      "External id": 34,"Record function id": 0, "Ev Idx": 33
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400662.631, "dur": 2.390,
    "args": {
      "External id": 35,"Record function id": 0, "Ev Idx": 34
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400662.901, "dur": 1.970,
    "args": {
      "External id": 36,"Record function id": 0, "Ev Idx": 35
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400666.431, "dur": 2.170,
    "args": {
      "External id": 37,"Record function id": 0, "Ev Idx": 36
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400666.661, "dur": 1.770,
    "args": {
      "External id": 38,"Record function id": 0, "Ev Idx": 37
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400669.761, "dur": 16.310,
    "args": {
      "External id": 39,"Record function id": 0, "Ev Idx": 38
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400670.861, "dur": 14.420,
    "args": {
      "External id": 40,"Record function id": 0, "Ev Idx": 39
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400707.431, "dur": 1.030,
    "args": {
      "External id": 41,"Record function id": 0, "Ev Idx": 40
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400709.401, "dur": 6.901,
    "args": {
      "External id": 42,"Record function id": 0, "Ev Idx": 41
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400709.871, "dur": 6.281,
    "args": {
      "External id": 43,"Record function id": 0, "Ev Idx": 42
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400710.901, "dur": 2.431,
    "args": {
      "External id": 44,"Record function id": 0, "Ev Idx": 43
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400714.022, "dur": 1.910,
    "args": {
      "External id": 45,"Record function id": 0, "Ev Idx": 44
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400717.212, "dur": 0.120,
    "args": {
      "External id": 46,"Record function id": 0, "Ev Idx": 45
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400718.002, "dur": 0.120,
    "args": {
      "External id": 47,"Record function id": 0, "Ev Idx": 46
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400720.792, "dur": 0.240,
    "args": {
      "External id": 48,"Record function id": 0, "Ev Idx": 47
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400721.532, "dur": 3.390,
    "args": {
      "External id": 49,"Record function id": 0, "Ev Idx": 48
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400721.772, "dur": 3.030,
    "args": {
      "External id": 50,"Record function id": 0, "Ev Idx": 49
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400722.202, "dur": 1.510,
    "args": {
      "External id": 51,"Record function id": 0, "Ev Idx": 50
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400724.082, "dur": 0.550,
    "args": {
      "External id": 52,"Record function id": 0, "Ev Idx": 51
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400725.392, "dur": 0.040,
    "args": {
      "External id": 53,"Record function id": 0, "Ev Idx": 52
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400725.782, "dur": 0.030,
    "args": {
      "External id": 54,"Record function id": 0, "Ev Idx": 53
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#4", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400769.913, "dur": 158.363,
    "args": {
      "External id": 55,"Record function id": 0, "Ev Idx": 54
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "flash_kernel_fwd", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400776.633, "dur": 147.073,
    "args": {
      "External id": 56,"Record function id": 0, "Ev Idx": 55
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400789.443, "dur": 125.992,
    "args": {
      "External id": 57,"Record function id": 0, "Sequence number": 7, "Fwd thread id": 0, "Ev Idx": 56
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn3_cuda_8a730d9::_flash_attn_forward", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400801.453, "dur": 105.882,
    "args": {
      "External id": 58,"Record function id": 0, "Ev Idx": 57
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn3_cuda_8a730d9::fwd", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400826.304, "dur": 46.791,
    "args": {
      "External id": 59,"Record function id": 0, "Ev Idx": 58
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400830.694, "dur": 5.200,
    "args": {
      "External id": 60,"Record function id": 0, "Ev Idx": 59
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400831.284, "dur": 4.320,
    "args": {
      "External id": 61,"Record function id": 0, "Ev Idx": 60
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400837.434, "dur": 2.310,
    "args": {
      "External id": 62,"Record function id": 0, "Ev Idx": 61
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400837.664, "dur": 1.920,
    "args": {
      "External id": 63,"Record function id": 0, "Ev Idx": 62
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400841.314, "dur": 2.100,
    "args": {
      "External id": 64,"Record function id": 0, "Ev Idx": 63
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400841.564, "dur": 1.700,
    "args": {
      "External id": 65,"Record function id": 0, "Ev Idx": 64
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400844.614, "dur": 15.270,
    "args": {
      "External id": 66,"Record function id": 0, "Ev Idx": 65
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400845.604, "dur": 13.550,
    "args": {
      "External id": 67,"Record function id": 0, "Ev Idx": 66
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400879.965, "dur": 0.980,
    "args": {
      "External id": 68,"Record function id": 0, "Ev Idx": 67
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400881.855, "dur": 6.460,
    "args": {
      "External id": 69,"Record function id": 0, "Ev Idx": 68
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400882.325, "dur": 5.850,
    "args": {
      "External id": 70,"Record function id": 0, "Ev Idx": 69
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400883.295, "dur": 2.270,
    "args": {
      "External id": 71,"Record function id": 0, "Ev Idx": 70
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400886.145, "dur": 1.810,
    "args": {
      "External id": 72,"Record function id": 0, "Ev Idx": 71
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400888.985, "dur": 0.120,
    "args": {
      "External id": 73,"Record function id": 0, "Ev Idx": 72
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400889.675, "dur": 0.120,
    "args": {
      "External id": 74,"Record function id": 0, "Ev Idx": 73
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400892.515, "dur": 0.230,
    "args": {
      "External id": 75,"Record function id": 0, "Ev Idx": 74
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400893.205, "dur": 3.350,
    "args": {
      "External id": 76,"Record function id": 0, "Ev Idx": 75
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400893.435, "dur": 3.000,
    "args": {
      "External id": 77,"Record function id": 0, "Ev Idx": 76
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400893.905, "dur": 1.500,
    "args": {
      "External id": 78,"Record function id": 0, "Ev Idx": 77
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400895.745, "dur": 0.510,
    "args": {
      "External id": 79,"Record function id": 0, "Ev Idx": 78
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400897.005, "dur": 0.050,
    "args": {
      "External id": 80,"Record function id": 0, "Ev Idx": 79
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400897.375, "dur": 0.030,
    "args": {
      "External id": 81,"Record function id": 0, "Ev Idx": 80
    }
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1393548402660.123, "dur": 1.792,
    "args": {
      "External id": 13,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 94,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.007576,      "warps per SM": 0.030303,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 94, "pid": 0, "tid": 7, "ts": 1393548402660.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400460.707, "dur": 12.330,
    "args": {
      "External id": 13,      "cbid": 211, "correlation": 94
    }
  },
  {
    "ph": "s", "id": 94, "pid": 2768517, "tid": 2768517, "ts": 1393548400460.707,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400477.397, "dur": 0.480,
    "args": {
      "External id": 5,      "cbid": 406, "correlation": 102
    }
  },
  {
    "ph": "f", "id": 102, "pid": 2768517, "tid": 2768517, "ts": 1393548400477.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400478.757, "dur": 0.160,
    "args": {
      "External id": 5,      "cbid": 406, "correlation": 104
    }
  },
  {
    "ph": "f", "id": 104, "pid": 2768517, "tid": 2768517, "ts": 1393548400478.757,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400479.267, "dur": 0.130,
    "args": {
      "External id": 5,      "cbid": 406, "correlation": 106
    }
  },
  {
    "ph": "f", "id": 106, "pid": 2768517, "tid": 2768517, "ts": 1393548400479.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400479.697, "dur": 0.130,
    "args": {
      "External id": 5,      "cbid": 406, "correlation": 108
    }
  },
  {
    "ph": "f", "id": 108, "pid": 2768517, "tid": 2768517, "ts": 1393548400479.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400480.057, "dur": 0.130,
    "args": {
      "External id": 5,      "cbid": 406, "correlation": 110
    }
  },
  {
    "ph": "f", "id": 110, "pid": 2768517, "tid": 2768517, "ts": 1393548400480.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400480.597, "dur": 0.150,
    "args": {
      "External id": 5,      "cbid": 406, "correlation": 112
    }
  },
  {
    "ph": "f", "id": 112, "pid": 2768517, "tid": 2768517, "ts": 1393548400480.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400481.337, "dur": 0.720,
    "args": {
      "External id": 5,      "cbid": 273, "correlation": 114
    }
  },
  {
    "ph": "f", "id": 114, "pid": 2768517, "tid": 2768517, "ts": 1393548400481.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::device_kernel<flash::enable_sm90<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<192>, cute::C<128>, cute::C<64> >, 64, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<192>, cute::C<64>, cute::C<128> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 384, false, false, false, false>, flash::DynamicPersistentTileScheduler<384, 32, false, false, true> > > >(flash::enable_sm90<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<192>, cute::C<128>, cute::C<64> >, 64, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<192>, cute::C<64>, cute::C<128> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 384, false, false, false, false>, flash::DynamicPersistentTileScheduler<384, 32, false, false, true> > >::Params)", "pid": 0, "tid": 7,
    "ts": 1393548402662.939, "dur": 81.279,
    "args": {
      "External id": 5,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 115,      "registers per thread": 128,      "shared memory": 93184,      "blocks per SM": 1.000000,      "warps per SM": 16.000000,      "grid": [132, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 115, "pid": 0, "tid": 7, "ts": 1393548402662.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400482.387, "dur": 4.830,
    "args": {
      "External id": 5,      "cbid": 211, "correlation": 115
    }
  },
  {
    "ph": "s", "id": 115, "pid": 2768517, "tid": 2768517, "ts": 1393548400482.387,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1393548402745.467, "dur": 1.728,
    "args": {
      "External id": 40,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 160,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.007576,      "warps per SM": 0.030303,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 160, "pid": 0, "tid": 7, "ts": 1393548402745.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400674.241, "dur": 10.000,
    "args": {
      "External id": 40,      "cbid": 211, "correlation": 160
    }
  },
  {
    "ph": "s", "id": 160, "pid": 2768517, "tid": 2768517, "ts": 1393548400674.241,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400688.141, "dur": 0.450,
    "args": {
      "External id": 32,      "cbid": 406, "correlation": 168
    }
  },
  {
    "ph": "f", "id": 168, "pid": 2768517, "tid": 2768517, "ts": 1393548400688.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400689.451, "dur": 0.140,
    "args": {
      "External id": 32,      "cbid": 406, "correlation": 170
    }
  },
  {
    "ph": "f", "id": 170, "pid": 2768517, "tid": 2768517, "ts": 1393548400689.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400689.951, "dur": 0.130,
    "args": {
      "External id": 32,      "cbid": 406, "correlation": 172
    }
  },
  {
    "ph": "f", "id": 172, "pid": 2768517, "tid": 2768517, "ts": 1393548400689.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400690.431, "dur": 0.150,
    "args": {
      "External id": 32,      "cbid": 406, "correlation": 174
    }
  },
  {
    "ph": "f", "id": 174, "pid": 2768517, "tid": 2768517, "ts": 1393548400690.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400690.871, "dur": 0.130,
    "args": {
      "External id": 32,      "cbid": 406, "correlation": 176
    }
  },
  {
    "ph": "f", "id": 176, "pid": 2768517, "tid": 2768517, "ts": 1393548400690.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400691.351, "dur": 0.150,
    "args": {
      "External id": 32,      "cbid": 406, "correlation": 178
    }
  },
  {
    "ph": "f", "id": 178, "pid": 2768517, "tid": 2768517, "ts": 1393548400691.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400692.181, "dur": 0.660,
    "args": {
      "External id": 32,      "cbid": 273, "correlation": 180
    }
  },
  {
    "ph": "f", "id": 180, "pid": 2768517, "tid": 2768517, "ts": 1393548400692.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::device_kernel<flash::enable_sm90<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<192>, cute::C<128>, cute::C<64> >, 64, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<192>, cute::C<64>, cute::C<128> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 384, false, false, false, false>, flash::DynamicPersistentTileScheduler<384, 32, false, false, true> > > >(flash::enable_sm90<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<192>, cute::C<128>, cute::C<64> >, 64, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<192>, cute::C<64>, cute::C<128> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 384, false, false, false, false>, flash::DynamicPersistentTileScheduler<384, 32, false, false, true> > >::Params)", "pid": 0, "tid": 7,
    "ts": 1393548402748.251, "dur": 79.392,
    "args": {
      "External id": 32,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 181,      "registers per thread": 128,      "shared memory": 93184,      "blocks per SM": 1.000000,      "warps per SM": 16.000000,      "grid": [132, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 181, "pid": 0, "tid": 7, "ts": 1393548402748.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400693.161, "dur": 4.540,
    "args": {
      "External id": 32,      "cbid": 211, "correlation": 181
    }
  },
  {
    "ph": "s", "id": 181, "pid": 2768517, "tid": 2768517, "ts": 1393548400693.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, std::array<char*, 1ul> >(int, at::native::FillFunctor<int>, std::array<char*, 1ul>)", "pid": 0, "tid": 7,
    "ts": 1393548402828.763, "dur": 1.887,
    "args": {
      "External id": 67,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 226,      "registers per thread": 16,      "shared memory": 0,      "blocks per SM": 0.007576,      "warps per SM": 0.030303,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 226, "pid": 0, "tid": 7, "ts": 1393548402828.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400848.794, "dur": 9.390,
    "args": {
      "External id": 67,      "cbid": 211, "correlation": 226
    }
  },
  {
    "ph": "s", "id": 226, "pid": 2768517, "tid": 2768517, "ts": 1393548400848.794,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400861.674, "dur": 0.380,
    "args": {
      "External id": 59,      "cbid": 406, "correlation": 234
    }
  },
  {
    "ph": "f", "id": 234, "pid": 2768517, "tid": 2768517, "ts": 1393548400861.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400862.814, "dur": 0.130,
    "args": {
      "External id": 59,      "cbid": 406, "correlation": 236
    }
  },
  {
    "ph": "f", "id": 236, "pid": 2768517, "tid": 2768517, "ts": 1393548400862.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400863.294, "dur": 0.130,
    "args": {
      "External id": 59,      "cbid": 406, "correlation": 238
    }
  },
  {
    "ph": "f", "id": 238, "pid": 2768517, "tid": 2768517, "ts": 1393548400863.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400863.714, "dur": 0.130,
    "args": {
      "External id": 59,      "cbid": 406, "correlation": 240
    }
  },
  {
    "ph": "f", "id": 240, "pid": 2768517, "tid": 2768517, "ts": 1393548400863.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400864.134, "dur": 0.130,
    "args": {
      "External id": 59,      "cbid": 406, "correlation": 242
    }
  },
  {
    "ph": "f", "id": 242, "pid": 2768517, "tid": 2768517, "ts": 1393548400864.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400864.644, "dur": 0.130,
    "args": {
      "External id": 59,      "cbid": 406, "correlation": 244
    }
  },
  {
    "ph": "f", "id": 244, "pid": 2768517, "tid": 2768517, "ts": 1393548400864.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400865.374, "dur": 0.590,
    "args": {
      "External id": 59,      "cbid": 273, "correlation": 246
    }
  },
  {
    "ph": "f", "id": 246, "pid": 2768517, "tid": 2768517, "ts": 1393548400865.374,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void cutlass::device_kernel<flash::enable_sm90<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<192>, cute::C<128>, cute::C<64> >, 64, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<192>, cute::C<64>, cute::C<128> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 384, false, false, false, false>, flash::DynamicPersistentTileScheduler<384, 32, false, false, true> > > >(flash::enable_sm90<flash::FlashAttnFwdSm90<flash::CollectiveMainloopFwdSm90<2, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cute::tuple<cute::C<192>, cute::C<128>, cute::C<64> >, 64, cutlass::bfloat16_t, float, cutlass::arch::Sm90, true, false, false, false, false, false, false, true, true, false, false, false>, flash::CollectiveEpilogueFwd<cute::tuple<cute::C<192>, cute::C<64>, cute::C<128> >, cute::tuple<cute::C<1>, cute::C<1>, cute::C<1> >, cutlass::bfloat16_t, cutlass::arch::Sm90, 384, false, false, false, false>, flash::DynamicPersistentTileScheduler<384, 32, false, false, true> > >::Params)", "pid": 0, "tid": 7,
    "ts": 1393548402831.771, "dur": 79.007,
    "args": {
      "External id": 59,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 247,      "registers per thread": 128,      "shared memory": 93184,      "blocks per SM": 1.000000,      "warps per SM": 16.000000,      "grid": [132, 1, 1],      "block": [512, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 247, "pid": 0, "tid": 7, "ts": 1393548402831.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400866.254, "dur": 4.351,
    "args": {
      "External id": 59,      "cbid": 211, "correlation": 247
    }
  },
  {
    "ph": "s", "id": 247, "pid": 2768517, "tid": 2768517, "ts": 1393548400866.254,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2768517, "tid": 2768517,
    "ts": 1393548400967.186, "dur": 1947.817,
    "args": {
            "cbid": 165, "correlation": 267
    }
  },
  {
    "ph": "s", "id": 267, "pid": 2768517, "tid": 2768517, "ts": 1393548400967.186,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "flash_kernel_fwd", "pid": 0, "tid": 7,
    "ts": 1393548402828.762, "dur": 82.017,
    "args": {
      "External id": 56
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "flash_kernel_fwd", "pid": 0, "tid": 7,
    "ts": 1393548402745.466, "dur": 82.178,
    "args": {
      "External id": 29
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "flash_kernel_fwd", "pid": 0, "tid": 7,
    "ts": 1393548402660.122, "dur": 84.097,
    "args": {
      "External id": 2
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 2768517, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 2768517, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 2768517, "tid": 0,
    "args": {
      "sort_index": 2768517
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 0, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 1, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 2, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 3, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 4, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 5, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 6, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 7, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 8, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 9, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 10, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 11, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 12, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 13, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 14, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393548400235.555, "pid": 15, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393548400235.555, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1393548400235.555, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1393548400235.555, "pid": 2768517, "tid": 2768517,
    "args": {
      "name": "thread 2768517 (python3)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 2768517, "tid": 2768517,
    "args": {
      "sort_index": 2768517
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1393548400235.555, "pid": 2768517, "tid": 2768517,
    "args": {
      "name": "thread 2768517 (python3)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1393548400235.555, "pid": 2768517, "tid": 2768517,
    "args": {
      "sort_index": 2768517
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 1393548400184.024, "dur": 2740.242,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393548400184.024,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 1393548400184.024
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 1393548403082.099
  }
  ],
  "traceName": "./traces/04_d_flash_kernels/8_16_1024_64_flashattn3.json"
}