633 lines
22 KiB
JSON
633 lines
22 KiB
JSON
|
|
{
|
|
"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": "F354D8A1E5E64E6BAA8B1A80A956C233",
|
|
"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": 4656971064223.358, "dur": 2259.639,
|
|
"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": [[292264448, 10438016, 128, 1], [292264448, 10438016, 128, 1], [292264448, 10438016, 128, 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": 4656971064256.665, "dur": 2219.828,
|
|
"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": [[292264448, 10438016, 128, 1], [292264448, 10438016, 128, 1], [292264448, 10438016, 128, 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": 4656971064269.951, "dur": 25.519,
|
|
"args": {
|
|
"External id": 515,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[292264448, 10438016, 128, 1], [], []], "Input Dims": [[1, 28, 81547, 128], [], []], "Ev Idx": 2
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971064286.317, "dur": 5.998,
|
|
"args": {
|
|
"External id": 516,"Record function id": 0, "Concrete Inputs": ["", "[1, 81547, 28, 128]", "[292264448, 128, 10438016, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[292264448, 10438016, 128, 1], [], [], []], "Input Dims": [[1, 28, 81547, 128], [], [], []], "Ev Idx": 3
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971064296.302, "dur": 1.974,
|
|
"args": {
|
|
"External id": 517,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[292264448, 10438016, 128, 1], [], []], "Input Dims": [[1, 28, 81547, 128], [], []], "Ev Idx": 4
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971064297.444, "dur": 0.547,
|
|
"args": {
|
|
"External id": 518,"Record function id": 0, "Concrete Inputs": ["", "[1, 81547, 28, 128]", "[292264448, 128, 10438016, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[292264448, 10438016, 128, 1], [], [], []], "Input Dims": [[1, 28, 81547, 128], [], [], []], "Ev Idx": 5
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971064299.035, "dur": 1.432,
|
|
"args": {
|
|
"External id": 519,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[292264448, 10438016, 128, 1], [], []], "Input Dims": [[1, 28, 81547, 128], [], []], "Ev Idx": 6
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971064299.906, "dur": 0.282,
|
|
"args": {
|
|
"External id": 520,"Record function id": 0, "Concrete Inputs": ["", "[1, 81547, 28, 128]", "[292264448, 128, 10438016, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[292264448, 10438016, 128, 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": 4656971064306.648, "dur": 2159.730,
|
|
"args": {
|
|
"External id": 521,"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": [[292264448, 128, 10438016, 1], [292264448, 128, 10438016, 1], [292264448, 128, 10438016, 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": 4656971064316.207, "dur": 1904.046,
|
|
"args": {
|
|
"External id": 522,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[292264448, 128, 10438016, 1], [], [], [], [], []], "Input Dims": [[1, 81547, 28, 128], [], [], [], [], []], "Ev Idx": 9
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971064323.468, "dur": 1893.873,
|
|
"args": {
|
|
"External id": 523,"Record function id": 0, "Concrete Inputs": ["[1, 81547, 28, 128]", "[292264448, 128, 10438016, 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": 4656971066229.956, "dur": 9.417,
|
|
"args": {
|
|
"External id": 524,"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": 4656971066242.635, "dur": 141.849,
|
|
"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": 2815989, "tid": 2815989,
|
|
"ts": 4656971066386.311, "dur": 7.845,
|
|
"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": 2815989, "tid": 2815989,
|
|
"ts": 4656971066461.641, "dur": 2.813,
|
|
"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": 2815989, "tid": 2815989,
|
|
"ts": 4656971066469.023, "dur": 4.772,
|
|
"args": {
|
|
"External id": 528,"Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[292264448, 128, 10438016, 1], [], []], "Input Dims": [[1, 81547, 28, 128], [], []], "Ev Idx": 15
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066471.984, "dur": 1.380,
|
|
"args": {
|
|
"External id": 529,"Record function id": 0, "Concrete Inputs": ["", "[1, 28, 81547, 128]", "[292264448, 10438016, 128, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[292264448, 128, 10438016, 1], [], [], []], "Input Dims": [[1, 81547, 28, 128], [], [], []], "Ev Idx": 16
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "overhead", "name": "Activity Buffer Request", "pid": -1, "tid": 0,
|
|
"ts": 4656971064339.960, "dur": 1560.449
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971065908.900, "dur": 1.970,
|
|
"args": {
|
|
"External id": 523, "cbid": 317, "correlation": 2329
|
|
}
|
|
},
|
|
{
|
|
"ph": "f", "id": 2329, "pid": 2815989, "tid": 2815989, "ts": 4656971065908.900,
|
|
"cat": "ac2g", "name": "ac2g", "bp": "e"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971065911.382, "dur": 292.493,
|
|
"args": {
|
|
"External id": 523, "cbid": 20, "correlation": 2330
|
|
}
|
|
},
|
|
{
|
|
"ph": "f", "id": 2330, "pid": 2815989, "tid": 2815989, "ts": 4656971065911.382,
|
|
"cat": "ac2g", "name": "ac2g", "bp": "e"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066246.771, "dur": 0.875,
|
|
"args": {
|
|
"External id": 525, "cbid": 317, "correlation": 2348
|
|
}
|
|
},
|
|
{
|
|
"ph": "f", "id": 2348, "pid": 2815989, "tid": 2815989, "ts": 4656971066246.771,
|
|
"cat": "ac2g", "name": "ac2g", "bp": "e"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066247.985, "dur": 132.850,
|
|
"args": {
|
|
"External id": 525, "cbid": 20, "correlation": 2349
|
|
}
|
|
},
|
|
{
|
|
"ph": "f", "id": 2349, "pid": 2815989, "tid": 2815989, "ts": 4656971066247.985,
|
|
"cat": "ac2g", "name": "ac2g", "bp": "e"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066397.642, "dur": 2.475,
|
|
"args": {
|
|
"External id": 521, "cbid": 200, "correlation": 2363
|
|
}
|
|
},
|
|
{
|
|
"ph": "f", "id": 2363, "pid": 2815989, "tid": 2815989, "ts": 4656971066397.642,
|
|
"cat": "ac2g", "name": "ac2g", "bp": "e"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066400.324, "dur": 0.432,
|
|
"args": {
|
|
"External id": 521, "cbid": 200, "correlation": 2364
|
|
}
|
|
},
|
|
{
|
|
"ph": "f", "id": 2364, "pid": 2815989, "tid": 2815989, "ts": 4656971066400.324,
|
|
"cat": "ac2g", "name": "ac2g", "bp": "e"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066402.401, "dur": 10.227,
|
|
"args": {
|
|
"External id": 521, "cbid": 273, "correlation": 2365
|
|
}
|
|
},
|
|
{
|
|
"ph": "f", "id": 2365, "pid": 2815989, "tid": 2815989, "ts": 4656971066402.401,
|
|
"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": 4656971066460.067, "dur": 665911.283,
|
|
"args": {
|
|
"External id": 521, "queued": 0, "device": 0, "context": 1, "stream": 7, "correlation": 2367, "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": 2367, "pid": 0, "tid": 7, "ts": 4656971066460.067,
|
|
"cat": "ac2g", "name": "ac2g", "bp": "e"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066418.414, "dur": 39.961,
|
|
"args": {
|
|
"External id": 521, "cbid": 211, "correlation": 2367
|
|
}
|
|
},
|
|
{
|
|
"ph": "s", "id": 2367, "pid": 2815989, "tid": 2815989, "ts": 4656971066418.414,
|
|
"cat": "ac2g", "name": "ac2g"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971066557.289, "dur": 665818.659,
|
|
"args": {
|
|
"cbid": 165, "correlation": 2384
|
|
}
|
|
},
|
|
{
|
|
"ph": "s", "id": 2384, "pid": 2815989, "tid": 2815989, "ts": 4656971066557.289,
|
|
"cat": "ac2g", "name": "ac2g"
|
|
},
|
|
{
|
|
"ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 2815989, "tid": 2815989,
|
|
"ts": 4656971732405.496, "dur": 3.877,
|
|
"args": {
|
|
"cbid": 165, "correlation": 2388
|
|
}
|
|
},
|
|
{
|
|
"ph": "s", "id": 2388, "pid": 2815989, "tid": 2815989, "ts": 4656971732405.496,
|
|
"cat": "ac2g", "name": "ac2g"
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 2815989, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 2815989, "tid": 0,
|
|
"args": {
|
|
"labels": "CPU"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 2815989, "tid": 0,
|
|
"args": {
|
|
"sort_index": 2815989
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 0, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 0, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 0"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 0, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000000
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 1, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 1, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 1"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 1, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000001
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 2, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 2, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 2"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 2, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000002
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 3, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 3, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 3"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 3, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000003
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 4, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 4, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 4"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 4, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000004
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 5, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 5, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 5"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 5, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000005
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 6, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 6, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 6"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 6, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000006
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 7, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 7, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 7"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 7, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000007
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 8, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 8, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 8"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 8, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000008
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 9, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 9, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 9"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 9, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000009
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 10, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 10, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 10"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 10, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000010
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 11, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 11, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 11"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 11, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000011
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 12, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 12, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 12"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 12, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000012
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 13, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 13, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 13"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 13, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000013
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 14, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 14, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 14"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 14, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000014
|
|
}
|
|
},
|
|
{
|
|
"name": "process_name", "ph": "M", "ts": 4656971063927.022, "pid": 15, "tid": 0,
|
|
"args": {
|
|
"name": "python"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_labels", "ph": "M", "ts": 4656971063927.022, "pid": 15, "tid": 0,
|
|
"args": {
|
|
"labels": "GPU 15"
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 15, "tid": 0,
|
|
"args": {
|
|
"sort_index": 5000015
|
|
}
|
|
},
|
|
{
|
|
"name": "thread_name", "ph": "M", "ts": 4656971063927.022, "pid": 0, "tid": 7,
|
|
"args": {
|
|
"name": "stream 7 "
|
|
}
|
|
},
|
|
{
|
|
"name": "thread_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 0, "tid": 7,
|
|
"args": {
|
|
"sort_index": 7
|
|
}
|
|
},
|
|
{
|
|
"name": "thread_name", "ph": "M", "ts": 4656971063927.022, "pid": 2815989, "tid": 2815989,
|
|
"args": {
|
|
"name": "thread 2815989 (python)"
|
|
}
|
|
},
|
|
{
|
|
"name": "thread_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 2815989, "tid": 2815989,
|
|
"args": {
|
|
"sort_index": 2815989
|
|
}
|
|
},
|
|
{
|
|
"name": "thread_name", "ph": "M", "ts": 4656971063927.022, "pid": 2815989, "tid": 2815989,
|
|
"args": {
|
|
"name": "thread 2815989 (python)"
|
|
}
|
|
},
|
|
{
|
|
"name": "thread_sort_index", "ph": "M", "ts": 4656971063927.022, "pid": 2815989, "tid": 2815989,
|
|
"args": {
|
|
"sort_index": 2815989
|
|
}
|
|
},
|
|
{
|
|
"ph": "X", "cat": "Trace", "ts": 4656971063875.249, "dur": 668543.486,
|
|
"pid": "Spans", "tid": "PyTorch Profiler",
|
|
"name": "PyTorch Profiler (0)",
|
|
"args": {
|
|
"Op count": 0
|
|
}
|
|
},
|
|
{
|
|
"name": "process_sort_index", "ph": "M", "ts": 4656971063875.249,
|
|
"pid": "Spans", "tid": 0,
|
|
"args": {
|
|
"sort_index": 536870912
|
|
}
|
|
},
|
|
{
|
|
"name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
|
|
"pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 4656971063875.249
|
|
},
|
|
{
|
|
"name": "Record Window End", "ph": "i", "s": "g",
|
|
"pid": "", "tid": "", "ts": 4656971732647.495
|
|
}
|
|
],
|
|
"traceName": "/data/wxy/profile_results/h3-sdpa-kernel-matrix-20260825-run1/jobs/gpu4-RVA-flash-contiguous.trace.json"
|
|
} |