{
    "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": "629D47721ADE408A98A5590182DD427E",
      "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": 2815989, "tid": 2815989,
    "ts": 4656948370549.062, "dur": 4753.455,
    "args": {
      "External id": 1,"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": [[876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [], [], [], [], []], "Input Dims": [[1, 28, 81547, 128], [1, 28, 81547, 128], [1, 28, 81547, 128], [], [], [], [], []], "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_scaled_dot_product_flash_attention", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370637.198, "dur": 4652.716,
    "args": {
      "External id": 2,"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": [[876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [876793344, 128, 10752, 1], [], [], [], []], "Input Dims": [[1, 28, 81547, 128], [1, 28, 81547, 128], [1, 28, 81547, 128], [], [], [], []], "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370671.400, "dur": 41.244,
    "args": {
      "External id": 3,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[876793344, 128, 10752, 1], [], []], "Input Dims": [[1, 28, 81547, 128], [], []], "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370696.844, "dur": 11.631,
    "args": {
      "External id": 4,"Record function id": 0, "Concrete Inputs": ["", "[1, 81547, 28, 128]", "[876793344, 10752, 128, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[876793344, 128, 10752, 1], [], [], []], "Input Dims": [[1, 28, 81547, 128], [], [], []], "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370715.280, "dur": 4.606,
    "args": {
      "External id": 5,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[876793344, 128, 10752, 1], [], []], "Input Dims": [[1, 28, 81547, 128], [], []], "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370718.203, "dur": 0.974,
    "args": {
      "External id": 6,"Record function id": 0, "Concrete Inputs": ["", "[1, 81547, 28, 128]", "[876793344, 10752, 128, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[876793344, 128, 10752, 1], [], [], []], "Input Dims": [[1, 28, 81547, 128], [], [], []], "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370721.720, "dur": 3.517,
    "args": {
      "External id": 7,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[876793344, 128, 10752, 1], [], []], "Input Dims": [[1, 28, 81547, 128], [], []], "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370723.855, "dur": 0.707,
    "args": {
      "External id": 8,"Record function id": 0, "Concrete Inputs": ["", "[1, 81547, 28, 128]", "[876793344, 10752, 128, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[876793344, 128, 10752, 1], [], [], []], "Input Dims": [[1, 28, 81547, 128], [], [], []], "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_flash_attention_forward", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370745.998, "dur": 4521.366,
    "args": {
      "External id": 9,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "81547", "81547", "0.", "False", "False", "0.088388347648318433", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "", "", ""], "Input Strides": [[876793344, 10752, 128, 1], [876793344, 10752, 128, 1], [876793344, 10752, 128, 1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[1, 81547, 28, 128], [1, 81547, 28, 128], [1, 81547, 28, 128], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370785.864, "dur": 4094.210,
    "args": {
      "External id": 10,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[876793344, 10752, 128, 1], [], [], [], [], []], "Input Dims": [[1, 81547, 28, 128], [], [], [], [], []], "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2815989, "tid": 2815989,
    "ts": 4656948370810.929, "dur": 4063.339,
    "args": {
      "External id": 11,"Record function id": 0, "Concrete Inputs": ["[1, 81547, 28, 128]", "[292264448, 3584, 128, 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": 2815989, "tid": 2815989,
    "ts": 4656948374899.052, "dur": 21.987,
    "args": {
      "External id": 12,"Record function id": 0, "Concrete Inputs": ["[1, 28, 81547]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 2815989, "tid": 2815989,
    "ts": 4656948374928.477, "dur": 200.234,
    "args": {
      "External id": 13,"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": 2815989, "tid": 2815989,
    "ts": 4656948375132.845, "dur": 8.902,
    "args": {
      "External id": 14,"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": 2815989, "tid": 2815989,
    "ts": 4656948375256.548, "dur": 6.989,
    "args": {
      "External id": 15,"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": 2815989, "tid": 2815989,
    "ts": 4656948375272.335, "dur": 10.768,
    "args": {
      "External id": 16,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[292264448, 3584, 128, 1], [], []], "Input Dims": [[1, 81547, 28, 128], [], []], "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
    "ts": 4656948375278.031, "dur": 4.070,
    "args": {
      "External id": 17,"Record function id": 0, "Concrete Inputs": ["", "[1, 28, 81547, 128]", "[292264448, 128, 3584, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[292264448, 3584, 128, 1], [], [], []], "Input Dims": [[1, 81547, 28, 128], [], [], []], "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "overhead", "name": "Activity Buffer Request", "pid": -1, "tid": 0,
    "ts": 4656948370840.902, "dur": 3598.708
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815989, "tid": 2815989,
    "ts": 4656948374455.851, "dur": 12.062,
    "args": {
      "External id": 11,      "cbid": 317, "correlation": 13
    }
  },
  {
    "ph": "f", "id": 13, "pid": 2815989, "tid": 2815989, "ts": 4656948374455.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815989, "tid": 2815989,
    "ts": 4656948374469.751, "dur": 373.895,
    "args": {
      "External id": 11,      "cbid": 20, "correlation": 14
    }
  },
  {
    "ph": "f", "id": 14, "pid": 2815989, "tid": 2815989, "ts": 4656948374469.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815989, "tid": 2815989,
    "ts": 4656948374937.183, "dur": 1.823,
    "args": {
      "External id": 13,      "cbid": 317, "correlation": 32
    }
  },
  {
    "ph": "f", "id": 32, "pid": 2815989, "tid": 2815989, "ts": 4656948374937.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815989, "tid": 2815989,
    "ts": 4656948374939.629, "dur": 181.346,
    "args": {
      "External id": 13,      "cbid": 20, "correlation": 33
    }
  },
  {
    "ph": "f", "id": 33, "pid": 2815989, "tid": 2815989, "ts": 4656948374939.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2815989, "tid": 2815989,
    "ts": 4656948375146.976, "dur": 3.075,
    "args": {
      "External id": 9,      "cbid": 200, "correlation": 47
    }
  },
  {
    "ph": "f", "id": 47, "pid": 2815989, "tid": 2815989, "ts": 4656948375146.976,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2815989, "tid": 2815989,
    "ts": 4656948375150.451, "dur": 0.730,
    "args": {
      "External id": 9,      "cbid": 200, "correlation": 48
    }
  },
  {
    "ph": "f", "id": 48, "pid": 2815989, "tid": 2815989, "ts": 4656948375150.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2815989, "tid": 2815989,
    "ts": 4656948375153.069, "dur": 20.805,
    "args": {
      "External id": 9,      "cbid": 273, "correlation": 49
    }
  },
  {
    "ph": "f", "id": 49, "pid": 2815989, "tid": 2815989, "ts": 4656948375153.069,
    "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": 4656948375244.360, "dur": 664714.517,
    "args": {
      "External id": 9,      "queued": 0, "device": 0, "context": 1,      "stream": 7, "correlation": 51,      "registers per thread": 255,      "shared memory": 65536,      "blocks per SM": 114.512817,      "warps per SM": 458.051270,      "grid": [638, 1, 28],      "block": [128, 1, 1],      "est. achieved occupancy %": 0
    }
  },
  {
    "ph": "f", "id": 51, "pid": 0, "tid": 7, "ts": 4656948375244.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2815989, "tid": 2815989,
    "ts": 4656948375185.706, "dur": 63.294,
    "args": {
      "External id": 9,      "cbid": 211, "correlation": 51
    }
  },
  {
    "ph": "s", "id": 51, "pid": 2815989, "tid": 2815989, "ts": 4656948375185.706,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815989, "tid": 2815989,
    "ts": 4656948375441.319, "dur": 664523.502,
    "args": {
            "cbid": 165, "correlation": 68
    }
  },
  {
    "ph": "s", "id": 68, "pid": 2815989, "tid": 2815989, "ts": 4656948375441.319,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815989, "tid": 2815989,
    "ts": 4656949039991.950, "dur": 3.959,
    "args": {
            "cbid": 165, "correlation": 72
    }
  },
  {
    "ph": "s", "id": 72, "pid": 2815989, "tid": 2815989, "ts": 4656949039991.950,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 2815989, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 2815989, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 2815989, "tid": 0,
    "args": {
      "sort_index": 2815989
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 4656948370304.939, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 4656948370304.939, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656948370304.939, "pid": 0, "tid": 7,
    "args": {
      "name": "stream 7 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 0, "tid": 7,
    "args": {
      "sort_index": 7
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656948370304.939, "pid": 2815989, "tid": 2815989,
    "args": {
      "name": "thread 2815989 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 2815989, "tid": 2815989,
    "args": {
      "sort_index": 2815989
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 4656948370304.939, "pid": 2815989, "tid": 2815989,
    "args": {
      "name": "thread 2815989 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 4656948370304.939, "pid": 2815989, "tid": 2815989,
    "args": {
      "sort_index": 2815989
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 4656948370194.652, "dur": 669811.167,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 4656948370194.652,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 4656948370194.652
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 4656949040222.426
  }
  ],
  "traceName": "/data/wxy/profile_results/h3-sdpa-kernel-matrix-20260825-run1/jobs/gpu4-RVA-flash-model_strided.trace.json"
}