{
    "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": "1C1EFFFD1ACB402E9123BFAFE94D20FA",
      "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": 2753042, "tid": 2753042,
    "ts": 1393423768585.163, "dur": 192.913,
    "args": {
      "External id": 1,"Record function id": 0, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "flash_kernel_fwd", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768598.783, "dur": 172.623,
    "args": {
      "External id": 2,"Record function id": 0, "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768619.133, "dur": 142.583,
    "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_attn2_cuda_8a730d9::_flash_attn_forward", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768637.644, "dur": 103.211,
    "args": {
      "External id": 4,"Record function id": 0, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn2_cuda_8a730d9::fwd", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768666.384, "dur": 61.541,
    "args": {
      "External id": 5,"Record function id": 0, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768683.504, "dur": 17.701,
    "args": {
      "External id": 6,"Record function id": 0, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768693.275, "dur": 6.090,
    "args": {
      "External id": 7,"Record function id": 0, "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768702.535, "dur": 2.900,
    "args": {
      "External id": 8,"Record function id": 0, "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768705.915, "dur": 1.430,
    "args": {
      "External id": 9,"Record function id": 0, "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768708.835, "dur": 3.970,
    "args": {
      "External id": 10,"Record function id": 0, "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768751.626, "dur": 3.040,
    "args": {
      "External id": 11,"Record function id": 0, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#3", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768791.316, "dur": 123.663,
    "args": {
      "External id": 12,"Record function id": 0, "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "flash_kernel_fwd", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768800.517, "dur": 109.482,
    "args": {
      "External id": 13,"Record function id": 0, "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768813.777, "dur": 88.552,
    "args": {
      "External id": 14,"Record function id": 0, "Sequence number": 6, "Fwd thread id": 0, "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn2_cuda_8a730d9::_flash_attn_forward", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768826.637, "dur": 63.131,
    "args": {
      "External id": 15,"Record function id": 0, "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn2_cuda_8a730d9::fwd", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768848.238, "dur": 31.030,
    "args": {
      "External id": 16,"Record function id": 0, "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768851.378, "dur": 6.000,
    "args": {
      "External id": 17,"Record function id": 0, "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768852.368, "dur": 4.580,
    "args": {
      "External id": 18,"Record function id": 0, "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768858.108, "dur": 2.490,
    "args": {
      "External id": 19,"Record function id": 0, "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768861.038, "dur": 1.390,
    "args": {
      "External id": 20,"Record function id": 0, "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768863.758, "dur": 1.990,
    "args": {
      "External id": 21,"Record function id": 0, "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768894.198, "dur": 2.410,
    "args": {
      "External id": 22,"Record function id": 0, "Ev Idx": 21
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#4", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768926.009, "dur": 110.312,
    "args": {
      "External id": 23,"Record function id": 0, "Ev Idx": 22
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "flash_kernel_fwd", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768934.429, "dur": 97.062,
    "args": {
      "External id": 24,"Record function id": 0, "Ev Idx": 23
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768946.179, "dur": 78.172,
    "args": {
      "External id": 25,"Record function id": 0, "Sequence number": 7, "Fwd thread id": 0, "Ev Idx": 24
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn2_cuda_8a730d9::_flash_attn_forward", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768956.540, "dur": 56.111,
    "args": {
      "External id": 26,"Record function id": 0, "Ev Idx": 25
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_flash_attn2_cuda_8a730d9::fwd", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768974.490, "dur": 28.560,
    "args": {
      "External id": 27,"Record function id": 0, "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768977.280, "dur": 5.140,
    "args": {
      "External id": 28,"Record function id": 0, "Ev Idx": 27
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768977.930, "dur": 4.080,
    "args": {
      "External id": 29,"Record function id": 0, "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768983.070, "dur": 2.330,
    "args": {
      "External id": 30,"Record function id": 0, "Ev Idx": 29
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768985.830, "dur": 1.500,
    "args": {
      "External id": 31,"Record function id": 0, "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768988.530, "dur": 2.020,
    "args": {
      "External id": 32,"Record function id": 0, "Ev Idx": 31
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 2753042, "tid": 2753042,
    "ts": 1393423769017.181, "dur": 1.980,
    "args": {
      "External id": 33,"Record function id": 0, "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "kernel", "name": "void flash::flash_fwd_kernel<Flash_fwd_kernel_traits<64, 128, 128, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<64, 128, 128, 4, cutlass::bfloat16_t> >, false, true, false, false, true, true, false, false>(flash::Flash_fwd_params)", "pid": 0, "tid": 7,
    "ts": 1393423768529.987, "dur": 121.281,
    "args": {
            "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 37,      "registers per thread": 255,      "shared memory": 49152,      "blocks per SM": 7.757576,      "warps per SM": 31.030304,      "grid": [8, 8, 16],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 37, "pid": 0, "tid": 7, "ts": 1393423768529.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768669.074, "dur": 0.390,
    "args": {
      "External id": 5,      "cbid": 200, "correlation": 46
    }
  },
  {
    "ph": "f", "id": 46, "pid": 2753042, "tid": 2753042, "ts": 1393423768669.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768669.594, "dur": 0.130,
    "args": {
      "External id": 5,      "cbid": 200, "correlation": 47
    }
  },
  {
    "ph": "f", "id": 47, "pid": 2753042, "tid": 2753042, "ts": 1393423768669.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768708.275, "dur": 0.200,
    "args": {
      "External id": 5,      "cbid": 200, "correlation": 70
    }
  },
  {
    "ph": "f", "id": 70, "pid": 2753042, "tid": 2753042, "ts": 1393423768708.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768713.715, "dur": 1.530,
    "args": {
      "External id": 5,      "cbid": 273, "correlation": 79
    }
  },
  {
    "ph": "f", "id": 79, "pid": 2753042, "tid": 2753042, "ts": 1393423768713.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void flash::flash_fwd_kernel<Flash_fwd_kernel_traits<64, 128, 128, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<64, 128, 128, 4, cutlass::bfloat16_t> >, false, true, false, false, true, true, false, false>(flash::Flash_fwd_params)", "pid": 0, "tid": 7,
    "ts": 1393423771124.293, "dur": 122.944,
    "args": {
      "External id": 5,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 80,      "registers per thread": 255,      "shared memory": 49152,      "blocks per SM": 7.757576,      "warps per SM": 31.030304,      "grid": [8, 8, 16],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 80, "pid": 0, "tid": 7, "ts": 1393423771124.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768715.775, "dur": 10.410,
    "args": {
      "External id": 5,      "cbid": 211, "correlation": 80
    }
  },
  {
    "ph": "s", "id": 80, "pid": 2753042, "tid": 2753042, "ts": 1393423768715.775,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768850.048, "dur": 0.350,
    "args": {
      "External id": 16,      "cbid": 200, "correlation": 89
    }
  },
  {
    "ph": "f", "id": 89, "pid": 2753042, "tid": 2753042, "ts": 1393423768850.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768850.508, "dur": 0.110,
    "args": {
      "External id": 16,      "cbid": 200, "correlation": 90
    }
  },
  {
    "ph": "f", "id": 90, "pid": 2753042, "tid": 2753042, "ts": 1393423768850.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768863.218, "dur": 0.190,
    "args": {
      "External id": 16,      "cbid": 200, "correlation": 113
    }
  },
  {
    "ph": "f", "id": 113, "pid": 2753042, "tid": 2753042, "ts": 1393423768863.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768866.588, "dur": 1.300,
    "args": {
      "External id": 16,      "cbid": 273, "correlation": 122
    }
  },
  {
    "ph": "f", "id": 122, "pid": 2753042, "tid": 2753042, "ts": 1393423768866.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void flash::flash_fwd_kernel<Flash_fwd_kernel_traits<64, 128, 128, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<64, 128, 128, 4, cutlass::bfloat16_t> >, false, true, false, false, true, true, false, false>(flash::Flash_fwd_params)", "pid": 0, "tid": 7,
    "ts": 1393423771248.357, "dur": 116.512,
    "args": {
      "External id": 16,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 123,      "registers per thread": 255,      "shared memory": 49152,      "blocks per SM": 7.757576,      "warps per SM": 31.030304,      "grid": [8, 8, 16],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 123, "pid": 0, "tid": 7, "ts": 1393423771248.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768868.288, "dur": 9.360,
    "args": {
      "External id": 16,      "cbid": 211, "correlation": 123
    }
  },
  {
    "ph": "s", "id": 123, "pid": 2753042, "tid": 2753042, "ts": 1393423768868.288,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768976.020, "dur": 0.350,
    "args": {
      "External id": 27,      "cbid": 200, "correlation": 132
    }
  },
  {
    "ph": "f", "id": 132, "pid": 2753042, "tid": 2753042, "ts": 1393423768976.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768976.460, "dur": 0.120,
    "args": {
      "External id": 27,      "cbid": 200, "correlation": 133
    }
  },
  {
    "ph": "f", "id": 133, "pid": 2753042, "tid": 2753042, "ts": 1393423768976.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768988.000, "dur": 0.200,
    "args": {
      "External id": 27,      "cbid": 200, "correlation": 156
    }
  },
  {
    "ph": "f", "id": 156, "pid": 2753042, "tid": 2753042, "ts": 1393423768988.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768991.330, "dur": 1.290,
    "args": {
      "External id": 27,      "cbid": 273, "correlation": 165
    }
  },
  {
    "ph": "f", "id": 165, "pid": 2753042, "tid": 2753042, "ts": 1393423768991.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void flash::flash_fwd_kernel<Flash_fwd_kernel_traits<64, 128, 128, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<64, 128, 128, 4, cutlass::bfloat16_t> >, false, true, false, false, true, true, false, false>(flash::Flash_fwd_params)", "pid": 0, "tid": 7,
    "ts": 1393423771366.149, "dur": 118.688,
    "args": {
      "External id": 27,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 166,      "registers per thread": 255,      "shared memory": 49152,      "blocks per SM": 7.757576,      "warps per SM": 31.030304,      "grid": [8, 8, 16],      "block": [128, 1, 1],      "est. achieved occupancy %": 13
    }
  },
  {
    "ph": "f", "id": 166, "pid": 0, "tid": 7, "ts": 1393423771366.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2753042, "tid": 2753042,
    "ts": 1393423768993.010, "dur": 8.530,
    "args": {
      "External id": 27,      "cbid": 211, "correlation": 166
    }
  },
  {
    "ph": "s", "id": 166, "pid": 2753042, "tid": 2753042, "ts": 1393423768993.010,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2753042, "tid": 2753042,
    "ts": 1393423769090.222, "dur": 2398.285,
    "args": {
            "cbid": 165, "correlation": 175
    }
  },
  {
    "ph": "s", "id": 175, "pid": 2753042, "tid": 2753042, "ts": 1393423769090.222,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "flash_kernel_fwd", "pid": 0, "tid": 7,
    "ts": 1393423771366.148, "dur": 118.690,
    "args": {
      "External id": 24
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "flash_kernel_fwd", "pid": 0, "tid": 7,
    "ts": 1393423771248.356, "dur": 116.514,
    "args": {
      "External id": 13
    }
  },
  {
    "ph": "X", "cat": "gpu_user_annotation", "name": "flash_kernel_fwd", "pid": 0, "tid": 7,
    "ts": 1393423771124.292, "dur": 122.946,
    "args": {
      "External id": 2
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 2753042, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 2753042, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 2753042, "tid": 0,
    "args": {
      "sort_index": 2753042
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 0, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 1, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 2, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 3, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 4, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 5, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 6, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 7, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 8, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 9, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 10, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 11, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 12, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 13, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 14, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1393423768481.145, "pid": 15, "tid": 0,
    "args": {
      "name": "python3"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1393423768481.145, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1393423768481.145, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1393423768481.145, "pid": 2753042, "tid": 2753042,
    "args": {
      "name": "thread 2753042 (python3)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 2753042, "tid": 2753042,
    "args": {
      "sort_index": 2753042
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1393423768481.145, "pid": 2753042, "tid": 2753042,
    "args": {
      "name": "thread 2753042 (python3)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1393423768481.145, "pid": 2753042, "tid": 2753042,
    "args": {
      "sort_index": 2753042
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 1393423768422.554, "dur": 3075.899,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1393423768422.554,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 1393423768422.554
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 1393423771635.965
  }
  ],
  "traceName": "./traces/04_d_flash_kernels/8_16_1024_64_flashattn2.json"
}