{
    "schemaVersion": 1,
    "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA RTX 6000D", "totalGlobalMem": 89171361792,
      "computeMajor": 12, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 1536,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 156
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 101376, "sharedMemPerMultiprocessor": 102400
    }
    ],
      "trace_id": "DE0091DAC29A4207946E2403F9D9C290",
      "record_shapes": 1,
      "cuda_driver_version": 13000,
      "cuda_runtime_version": 13000,
      "cupti_version": 130001,
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1782967788000000000,
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::scaled_dot_product_attention", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339753.340, "dur": 7600.358,
    "args": {
      "External id": 513,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "0.", "False", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [], [], [], [], []], "Input Dims": [[1, 28, 41778, 128], [1, 28, 41778, 128], [1, 28, 41778, 128], [], [], [], [], []], "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339789.910, "dur": 7554.021,
    "args": {
      "External id": 514,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "False", "False", "0.088388347648318433"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [149732352, 5347584, 128, 1], [], [], [], []], "Input Dims": [[1, 28, 41778, 128], [1, 28, 41778, 128], [1, 28, 41778, 128], [], [], [], []], "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339802.678, "dur": 22.400,
    "args": {
      "External id": 515,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[149732352, 5347584, 128, 1], [], []], "Input Dims": [[1, 28, 41778, 128], [], []], "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339816.376, "dur": 5.726,
    "args": {
      "External id": 516,"Record function id": 0, "Concrete Inputs": ["", "[1, 41778, 28, 128]", "[149732352, 128, 5347584, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[149732352, 5347584, 128, 1], [], [], []], "Input Dims": [[1, 28, 41778, 128], [], [], []], "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339826.211, "dur": 2.553,
    "args": {
      "External id": 517,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[149732352, 5347584, 128, 1], [], []], "Input Dims": [[1, 28, 41778, 128], [], []], "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339827.810, "dur": 0.637,
    "args": {
      "External id": 518,"Record function id": 0, "Concrete Inputs": ["", "[1, 41778, 28, 128]", "[149732352, 128, 5347584, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[149732352, 5347584, 128, 1], [], [], []], "Input Dims": [[1, 28, 41778, 128], [], [], []], "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339829.773, "dur": 2.086,
    "args": {
      "External id": 519,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[149732352, 5347584, 128, 1], [], []], "Input Dims": [[1, 28, 41778, 128], [], []], "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339831.066, "dur": 0.391,
    "args": {
      "External id": 520,"Record function id": 0, "Concrete Inputs": ["", "[1, 41778, 28, 128]", "[149732352, 128, 5347584, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[149732352, 5347584, 128, 1], [], [], []], "Input Dims": [[1, 28, 41778, 128], [], [], []], "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339836.528, "dur": 7484.879,
    "args": {
      "External id": 521,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "41778", "41778", "0.", "False", "False", "0.088388347648318433", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "", "", ""], "Input Strides": [[149732352, 128, 5347584, 1], [149732352, 128, 5347584, 1], [149732352, 128, 5347584, 1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[1, 41778, 28, 128], [1, 41778, 28, 128], [1, 41778, 28, 128], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339845.935, "dur": 4092.302,
    "args": {
      "External id": 522,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[149732352, 128, 5347584, 1], [], [], [], [], []], "Input Dims": [[1, 41778, 28, 128], [], [], [], [], []], "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2815985, "tid": 2815985,
    "ts": 4656938339852.237, "dur": 4081.074,
    "args": {
      "External id": 523,"Record function id": 0, "Concrete Inputs": ["[1, 41778, 28, 128]", "[149732352, 128, 5347584, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815985, "tid": 2815985,
    "ts": 4656938343952.191, "dur": 1658.243,
    "args": {
      "External id": 524,"Record function id": 0, "Concrete Inputs": ["[1, 28, 41778]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815985, "tid": 2815985,
    "ts": 4656938345616.682, "dur": 1612.704,
    "args": {
      "External id": 525,"Record function id": 0, "Concrete Inputs": ["[2]", "29", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347231.875, "dur": 6.550,
    "args": {
      "External id": 526,"Record function id": 0, "Concrete Inputs": ["[]", "29", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347313.760, "dur": 5.018,
    "args": {
      "External id": 527,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347326.372, "dur": 12.486,
    "args": {
      "External id": 528,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[149732352, 128, 5347584, 1], [], []], "Input Dims": [[1, 41778, 28, 128], [], []], "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347332.058, "dur": 6.148,
    "args": {
      "External id": 529,"Record function id": 0, "Concrete Inputs": ["", "[1, 28, 41778, 128]", "[149732352, 5347584, 128, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[149732352, 128, 5347584, 1], [], [], []], "Input Dims": [[1, 41778, 28, 128], [], [], []], "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "overhead", "name": "Activity Buffer Request", "pid": -1, "tid": 0,
    "ts": 4656938339868.672, "dur": 2725.729
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815985, "tid": 2815985,
    "ts": 4656938342608.575, "dur": 3.507,
    "args": {
      "External id": 523,      "cbid": 317, "correlation": 2334
    }
  },
  {
    "ph": "f", "id": 2334, "pid": 2815985, "tid": 2815985, "ts": 4656938342608.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815985, "tid": 2815985,
    "ts": 4656938342612.963, "dur": 1298.400,
    "args": {
      "External id": 523,      "cbid": 20, "correlation": 2335
    }
  },
  {
    "ph": "f", "id": 2335, "pid": 2815985, "tid": 2815985, "ts": 4656938342612.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815985, "tid": 2815985,
    "ts": 4656938343965.864, "dur": 1.425,
    "args": {
      "External id": 524,      "cbid": 317, "correlation": 2346
    }
  },
  {
    "ph": "f", "id": 2346, "pid": 2815985, "tid": 2815985, "ts": 4656938343965.864,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815985, "tid": 2815985,
    "ts": 4656938343967.817, "dur": 1636.239,
    "args": {
      "External id": 524,      "cbid": 20, "correlation": 2347
    }
  },
  {
    "ph": "f", "id": 2347, "pid": 2815985, "tid": 2815985, "ts": 4656938343967.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815985, "tid": 2815985,
    "ts": 4656938345622.556, "dur": 0.540,
    "args": {
      "External id": 525,      "cbid": 317, "correlation": 2357
    }
  },
  {
    "ph": "f", "id": 2357, "pid": 2815985, "tid": 2815985, "ts": 4656938345622.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815985, "tid": 2815985,
    "ts": 4656938345623.559, "dur": 1600.886,
    "args": {
      "External id": 525,      "cbid": 20, "correlation": 2358
    }
  },
  {
    "ph": "f", "id": 2358, "pid": 2815985, "tid": 2815985, "ts": 4656938345623.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347243.186, "dur": 1.518,
    "args": {
      "External id": 521,      "cbid": 200, "correlation": 2372
    }
  },
  {
    "ph": "f", "id": 2372, "pid": 2815985, "tid": 2815985, "ts": 4656938347243.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347245.190, "dur": 0.313,
    "args": {
      "External id": 521,      "cbid": 200, "correlation": 2373
    }
  },
  {
    "ph": "f", "id": 2373, "pid": 2815985, "tid": 2815985, "ts": 4656938347245.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347246.755, "dur": 5.770,
    "args": {
      "External id": 521,      "cbid": 273, "correlation": 2374
    }
  },
  {
    "ph": "f", "id": 2374, "pid": 2815985, "tid": 2815985, "ts": 4656938347246.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void pytorch_flash::flash_fwd_kernel<Flash_fwd_kernel_traits<128, 128, 64, 4, false, false, cutlass::bfloat16_t, Flash_kernel_traits<128, 128, 64, 4, cutlass::bfloat16_t> >, false, false, false, false, false, true, false, false>(pytorch_flash::Flash_fwd_params)", "pid": 0, "tid": 7,
    "ts": 4656938347307.623, "dur": 175158.471,
    "args": {
      "External id": 521,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 2376,      "registers per thread": 255,      "shared memory": 65536,      "blocks per SM": 58.692307,      "warps per SM": 234.769226,      "grid": [327, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 2376, "pid": 0, "tid": 7, "ts": 4656938347307.623,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347256.040, "dur": 53.157,
    "args": {
      "External id": 521,      "cbid": 211, "correlation": 2376
    }
  },
  {
    "ph": "s", "id": 2376, "pid": 2815985, "tid": 2815985, "ts": 4656938347256.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815985, "tid": 2815985,
    "ts": 4656938347475.645, "dur": 174998.260,
    "args": {
            "cbid": 165, "correlation": 2393
    }
  },
  {
    "ph": "s", "id": 2393, "pid": 2815985, "tid": 2815985, "ts": 4656938347475.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815985, "tid": 2815985,
    "ts": 4656938522512.354, "dur": 5.047,
    "args": {
            "cbid": 165, "correlation": 2397
    }
  },
  {
    "ph": "s", "id": 2397, "pid": 2815985, "tid": 2815985, "ts": 4656938522512.354,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 2815985, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 2815985, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 2815985, "tid": 0,
    "args": {
      "sort_index": 2815985
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656938339457.193, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656938339457.193, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656938339457.193, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656938339457.193, "pid": 2815985, "tid": 2815985,
    "args": {
      "name": "thread 2815985 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 2815985, "tid": 2815985,
    "args": {
      "sort_index": 2815985
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656938339457.193, "pid": 2815985, "tid": 2815985,
    "args": {
      "name": "thread 2815985 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656938339457.193, "pid": 2815985, "tid": 2815985,
    "args": {
      "sort_index": 2815985
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 4656938339391.509, "dur": 183140.508,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656938339391.509,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 4656938339391.509
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 4656938522850.590
  }
  ],
  "traceName": "/data/wxy/profile_results/h3-sdpa-kernel-matrix-20260825-run1/jobs/gpu0-F3-flash-contiguous.trace.json"
}